Kyruus Health wanted to improve patient-to-provider search, which was hindered by checkbox-based interfaces and the need for patients to translate symptoms into medical jargon. These traditional flows often returned irrelevant or incomplete provider results. The friction created decision fatigue and risked higher abandonment. Kyruus Health also needed an experience that matched expectations set by modern conversational GenAI tools. Kyruus Health partnered with Tribe to accelerate a generative AI transformation from proof of concept to general availability. Together they built “Guide,” a conversational search experience that let patients use natural language queries and received provider matches in seconds. The system routed queries through an API Gateway to a conversation service using Claude 3.5 Sonnet on Amazon Bedrock, extracted structured clinical parameters, retrieved candidates via Amazon OpenSearch, and generated clear conversational recommendations. Tribe also supported testing infrastructure, hallucination mitigation, and model validation while embedding with Kyruus Health’s engineering team. The partnership shortened the delivery timeline versus internal expectations. Instead of taking 4–6 months, the team reached production in two months. With Guide in production, Kyruus Health reported improved conversion for members scheduling appointments and faster, more reliable discovery of appropriate care options. The new experience also supported higher provider match satisfaction and faster appointment scheduling.

Project Details

Follett Software wanted to modernize its Destiny® Library Manager from a static catalog into an intelligent assistant for K-12 librarians. Librarians managed collections that sometimes exceeded 100,000+ titles using tools that could be complex, costly, and less intuitive. Although Follett had an existing AI team and strong ML foundations, it struggled to move beyond early LLM experimentation into a usable product experience. The company needed a clearer path to simplify the interface while delivering actionable insights that improved librarian workflows and ROI. A prototype AI-powered chat interface was built to let librarians interact with their collections using natural language. The solution used a custom text-to-SQL approach with multiple LLM layers (including Anthropic Claude models) to translate questions into reliable SQL queries. It followed a retrieval-augmented generation (RAG) architecture to return table results and support filtering and refinement. The system was designed to handle typos and understand library-specific concepts like the Dewey Decimal System, and it was implemented as a containerized cloud application on AWS Bedrock with an automated deployment pipeline. By the end of the engagement, Follett had a working prototype that enabled natural-language querying of school library collections. The interface returned clear answers and surfaced purchase and weeding recommendations, improving decision-making and creating a potential upsell pathway. Follett’s engineering team was also upskilled with patterns and frameworks to support and extend the solution internally. The work positioned Follett to scale AI as a platform-level capability across its broader portfolio of sellable products and suites.

Project Details

Keyloop’s aftersales operations relied on a manual, inefficient process for scheduling technicians. Workshop administrators built daily schedules from scratch despite constantly changing variables such as job types, technician skills, and bay availability. The lack of automation created administrative burden and constrained growth. It also negatively affected customer experience and left revenue unrealized. Keyloop partnered with Tribe AI to build an intelligent scheduling engine for its Service Hub product. The system incorporated real-world constraints like technician skills and job priority to automatically generate optimized daily schedules. Advisors started each day with a data-informed plan they could fine-tune rather than a blank slate. The approach included a technician schedule optimizer and a human-in-the-loop review process to support trust, accuracy, and compliance. The AI-driven scheduling capability reduced the need for workshop administrators to create schedules from scratch each morning. It improved workshop throughput and job prioritization by producing a ready-to-use daily schedule. The change supported higher technician utilization and faster service turnaround for dealerships. Keyloop also established a foundation to expand AI into additional initiatives, including work to accelerate OEM certifications through automation and structured schema mapping.

Project Details

Avalon managed over 60 complex, frequently updated policy documents used to determine patient eligibility for diagnostic tests. The prior authorization process was manual and time-consuming, requiring staff and providers to sift through extensive documentation. This created operational inefficiencies and increased the risk of delays or errors in patient care. Avalon partnered with Tribe to build a customized proof of concept using large language models to extract key information from policy documents and generate medically accurate prior authorization questions. The team used Claude 3 Opus to parse policy language, identify relevant coverage sections, and produce structured clinical questions for human reviewers. The PoC also supported PDF uploads, model selection, auto-generated test lists with manual adjustment, and a feedback loop to refine outputs with human-in-the-loop oversight. The PoC pilot achieved 100% precision in producing clinically accurate follow-up questions for human reviewers and delivered 83% recall across the policy documents tested, exceeding internal benchmarks. These results supported fewer eligibility assessment errors and faster review times. Avalon also shifted its policy documentation review cadence from annual to monthly cycles and planned to expand beyond the initial four test policies.

Project Details

Ataccama accelerated its use of AI but struggled to maintain and leverage quality data at scale. Department-level experiments and ad hoc automation created isolated efficiency gains while fragmenting the company’s trusted data foundation. The impact showed up in operations and go-to-market work, including 180 person-days per year lost to manual RFP responses and 80–150 hours per deal spent on repetitive sales preparation. Support operations also suffered, with 60–70% of tickets stuck in manual triage and information scattered across Slack, Notion, Jira, and Google Drive. Ataccama partnered with Tribe to operationalize data trust inside its own business through an AI Acceleration Framework. The framework applied governance, observability, and data quality principles to identify and automate high-impact repetitive work while preserving oversight. It implemented automated RFP responses and content generation using governed internal data, and it introduced Level 1 support triage powered by classification and enrichment models trained on trusted sources. An AI Program Office was also established to centralize governance, track ROI, and ensure consistent and compliant AI deployment across teams. In parallel, Ataccama extended its data trust foundation externally by creating an enterprise-grade Model Context Protocol (MCP) server. This transformed Ataccama ONE Agent’s 14+ capabilities into AI-ready, governed services with authentication, observability, and scalability for use across tools like Power BI, Claude Desktop, Snowflake Cortex, and developer IDEs. Internally, the company captured measurable gains by reducing manual effort and improving responsiveness in sales and support workflows. The dual approach positioned Ataccama as a cross-platform AI Trust Layer while enabling AI systems to query trusted enterprise data more directly.

Project Details

MyFitnessPal wanted to improve its user experience by incorporating generative AI into its product suite. The team needed additional capacity to scope, prototype, and iterate on innovative features. A key goal was increasing engagement by enabling users to log food conversationally instead of searching and typing each item. The desired experience required accurately capturing items, quantities, and even brands or packaged foods in a single step. MyFitnessPal brought in an external team to extend its product organization and help kickstart its AI product roadmap. The team built a functional demo in a sprint for a prototype called Voice Log that supported logging multiple items in one interaction. The flow let members enable the feature in-app, grant microphone permissions, speak meals in everyday language, and receive best-match suggestions from the database. The implementation used a full-stack, cloud-based approach integrated with MyFitnessPal’s existing environment. The project delivered a working Voice Log demo that streamlined multi-item food logging into a single conversational interaction. It also helped MyFitnessPal accelerate its AI product roadmap and informed how the company approached experimentation. Leadership was influenced to invest further in a culture of experimentation and to create a dedicated prototyping environment. MyFitnessPal positioned itself to identify and develop additional AI-powered features to strengthen its leadership in health and fitness.

Project Details

A global consulting firm specializing in corporate turnarounds and bankruptcy restructuring needed a faster way to analyze complex vendor ecosystems during diligence. The existing approach required weeks of manual effort to classify thousands of vendors into a multi-level taxonomy. Limited and inconsistent vendor information made the work error-prone and difficult to scale. The slow, manual process delayed downstream diligence deliverables and insight delivery. Tribe AI partnered with the firm to design and implement an AI-powered vendor classification engine inside the firm’s proprietary due diligence platform. The system used OpenAI GPT-4o and GPT-3.5-turbo with structured outputs, combined with historical classification patterns, deterministic business rules, and confidence scoring. It ingested raw vendor data, generated category predictions, applied rules-based mappings, and routed low-confidence cases for human review. A feedback loop captured human corrections to improve classifications over time. The firm categorized 18,000+ vendors in a live engagement with 97% accuracy, validated by subject matter experts. The new workflow reduced process time from 9–12 days to 45 minutes. The engine was used across the firm’s Private Equity and Turnaround & Restructuring service lines. It was actively deployed on hundreds of thousands of vendors and set as the standard going forward.

Skills

Project Details

A global consulting firm’s investment diligence process relied on “outside-in” research across public sources to assess risks, opportunities, and trends for target companies. The work was constrained by manual workflows, with analysts spending days or weeks gathering and stitching together fragmented signals from news, filings, job postings, and reviews. This created delays, risked missed signals, and led to inconsistencies across projects. The firm aimed to cut execution time down to 3 days by using an AI-driven approach. The firm partnered with Tribe AI to build an AI-powered research assistant embedded in its proprietary due diligence platform. The assistant aggregated data from public and proprietary sources and extracted relevant signals such as financials, employee rosters, salary data, leadership changes, and hiring trends. It presented insights in a searchable, curated dashboard with source citations and outputs compatible with Excel and PowerPoint. The system used NLP-based extraction, relevance ranking, and feedback loops to improve results over time. The implementation accelerated outside-in due diligence and improved consistency across projects. Outside-in research time was reduced from 7–10 days to under 1 day. Signal coverage increased by 30%, helping capture insights that manual workflows would have missed. Manual analyst hours were reduced by 40% per diligence project, and reporting was accelerated by 2–3 days to enable earlier decision-making.

Skills

Project Details

A hardware and software development company set out to build the industry’s first commercial utility Ground Penetrating Radar (GPR) system for safely verifying underground utilities before digging. Standard site-assessment procedures still produced frequent inaccuracies, causing unplanned disruptions, delays, and financial liability for contractors. The team needed higher accuracy and portability than existing offerings and believed AI could materially improve detection performance. They also wanted the work to support patentable intellectual property. An end-to-end prototype was built to analyze scans from a custom GPR device and detect and locate underground objects. The system used AWS Elastic Kubernetes Service (EKS) for workload management, Python for machine learning, and React/Node for the application interface and backend. IMU data was integrated to correct scans and add context about ground conditions, while automated ingestion and preprocessing (e.g., de-wow, normalization, denoising, contrast enhancement, and data translation) prepared data for model inference. A tablet-based UI allowed users to run inference locally and review results as points or bounding boxes with confidence scores, with depth estimation informed by user-specified substrate type. By the end of the initial engagement, the team received a highly accurate (~90%+) prototype that could detect and locate underground infrastructure from the custom GPR scans. The handheld device concept targeted quick field adoption with an expected learning curve of about five minutes and produced a scan for a 10’ x 10’ area in about three minutes. The radar capability extended up to 5 feet deep, and a visual map was generated within about 30 seconds after scan completion. The prototype’s preprocessing and system design contributed to a successful patent application and helped attract investor interest and new funding to move the product into production.

Project Details

TigerConnect faced persistent friction in clinical operations caused by inefficient on-call scheduling and shift swaps. Providers had to manually complete multiple steps to find who was on call or initiate a shift change, often wasting several minutes per inquiry. The delays compounded in time-sensitive care settings and hurt responsiveness and communication. With an estimated 300,000 to 400,000 end users across roughly 170 client organizations, the issue represented a widespread operational hurdle. TigerConnect also needed the initiative to succeed because it was positioned as the company’s first AI-driven product feature. TigerConnect partnered with Tribe to build a modular, LLM-powered scheduling assistant that handled on-call queries, contact lookups, and shift swaps through a natural language interface. The system used an agentic architecture where an LLM router detected user intent and routed requests to one of three specialized child agents (On-Call, Contact Info, Swap). The LLM dynamically orchestrated function calls to blend natural language understanding with deterministic backend actions like database lookups. The assistant was implemented as a proof of concept and validated against real-world interaction patterns with extensibility in mind. The LLM assistant reduced time spent by providers on on-call tasks by cutting steps required to complete common scheduling actions. It improved user experience by enabling natural language requests such as identifying the on-call provider or arranging a swap. The modular agent design demonstrated scalability across TigerConnect’s client base, supporting expansion from an initial small rollout to a broader set of organizations. The project also validated TigerConnect’s ability to integrate LLM capabilities into clinical tools, creating a foundation for additional AI features and monetization paths.

Project Details

Galley needed to turn an AI-forward vision into production-grade generative AI features. A major pain point was onboarding: getting customer recipe data into the system quickly and accurately. The internal team lacked specialized expertise to move from experimentation to a scalable, extensible implementation. They also needed the work to be strategically valuable beyond a single prototype. Tribe AI partnered with Galley starting with a rapid design sprint to define scope and build a working recipe-parser prototype. The team then developed an AI-powered Recipe Importer agent that ingested recipes from formats like PDFs and spreadsheets and mapped them into Galley’s internal schema using LLMs. The solution expanded into multiple components, including single-recipe and bulk parsing, menu plan importing, and ingredient normalization/deduplication. Tribe also implemented production infrastructure and an end-to-end testing framework with LLM-based “judges” to validate pipeline steps. After releasing the recipe importer in February, Galley reduced its time to close from about 90 days to 29 days by the end of Q1. Manual recipe input time dropped from up to 10 minutes per recipe to under 1 minute. The importer became a centerpiece of sales demos and contributed to measurable, immediate business impact. The project also established a foundation for broader agent-based AI initiatives across Galley’s product strategy.

Project Details

A leading enterprise tech company struggled to deliver consistent, high-quality sales enablement and coaching across global teams. Critical leadership guidance existed, but it was fragmented across decks, recordings, and internal memos. Live training and manual coaching did not scale well, making it hard to provide personalized, interactive support for sellers. The organization needed a way to operationalize institutional knowledge for broad, day-to-day use. The company partnered with Tribe AI to build an AI-powered coaching assistant modeled after a C-suite sales leader. The solution used existing leadership content—such as keynote talks, strategic documents, and deal guidance—to produce conversational, just-in-time coaching. It included a chat experience with text and voice input/output, a synthetic voice modeled after the leader, and an optional video/avatar layer for engagement. The platform was designed for enterprise use, including SSO integration and infrastructure intended to support large-scale usage. During internal testing and rollout, the assistant showed early signs of value for scalable coaching delivery. The organization expected it to accelerate seller ramp time, improve decision-making, and enhance customer conversations. It was also projected to shorten sales cycles and increase win rates by standardizing strategic guidance and improving deal execution. By reducing reliance on live coaching, the company expected to lower enablement costs and improve operational efficiency as adoption expanded.

Skills

Project Details

An insurance company wanted to use machine learning to turn existing data into a business asset that streamlined processes, produced insights, and drove revenue. The team had data and internal capability but lacked clarity on which ML use case would create the most business value. They also faced constraints from operating in a live pricing environment where distributors and price structures changed frequently. In addition, historical underwriting decisions kept price changes in a narrow band, limiting training data for larger pricing moves. Tribe ran a discovery effort to align stakeholders on the highest-value ML opportunity and ultimately focused the project on pricing optimization. The team then built a custom pricing algorithm and designed an intermediate experiment to gather the data needed to optimize premiums at scale. Because a true blind trial was not feasible, the experiment varied price changes by conversion likelihood to observe response across different price points. This approach created a path to learn price elasticity despite limited historical variation. The experiment produced an immediate 2.5% lift in premiums across the company. The project also delivered ROI in the first week, according to the product manager on the engagement. As the dynamic pricing model was rolled out and evaluated, the company estimated that expanding it from the test bucket to all policies could yield a 7–12% lift in premiums. The work also generated broader insights into the company’s book of business to support future optimization initiatives.

Skills

Project Details

Sonova wanted to mature and scale its machine learning function but faced bottlenecks in engineering productivity, model deployment, and operational reliability. Early models (including customer churn) existed, but automation, consistent infrastructure, and standardized production processes were missing. The 5-person data science team was overstretched and spent significant time on manual inference scoring and retraining. This slowed speed to market and increased risk around uptime, accuracy, and knowledge silos. To address this, a scalable ML infrastructure and repeatable blueprint were implemented to streamline deployments and reduce manual effort. Over 12 weeks, an MLOps framework was delivered to automate deployment, monitoring, and continuous delivery starting with the churn prediction model. Automated pipelines covered training, inference, evaluation, and reporting on schedules or triggered by production code changes. Testing/CI-CD practices, documentation, and real-time metrics for drift and underperformance detection were also put in place to support ongoing expansion. As a result, manual workflows were automated by up to ~32 hours/month per model, increasing engineering capacity. In 12 weeks, Sonova increased data science team capacity by 20% and reduced engineering overhead while improving production reliability. The effort also delivered a 20% reduction in engineering spend and fully operationalized the churn prediction model within the project timeline. Standardized practices reduced silos and improved continuity through shared documentation and enablement.

Skills

Project Details

Fantasmo needed to deploy its camera-based positioning algorithm in the field for a micromobility customer that required compliance enforcement without relying on cellular connectivity. The team wanted the full mapping and positioning workflow to run directly on customer hardware to avoid bandwidth and connection limitations. However, getting a large machine learning model to run on a low-power embedded chip required specialized expertise that would have taken time to hire. Fantasmo worked with Tribe to bring on Shalom, a machine learning engineer and researcher with prior experience deploying ML models on small devices. The team targeted a Luxonis chip to lower hardware costs while meeting power constraints. Shalom iterated on multiple approaches, built tooling to make the model trainable, and redesigned the model architecture to reduce compute and memory requirements. He also made the solution compatible with OpenVino to improve performance on the target platform. The project produced an infrastructure-less, connection-less positioning solution ready to be packaged and launched. The team reduced the model from 15MB to 0.6MB, significantly shrinking the on-device footprint. They also reduced compute costs and increased runtime speed versus the original implementation, helping Fantasmo move from feasibility concerns to a deployable embedded solution.

Skills

Project Details

Challenge: A software leader providing end-to-end compliance and audit management had stalled on an AI-driven security product after model accuracy plateaued around 70%. The internal team needed an NLP specialist to optimize the existing model but lacked in-house expertise. The manual nature of the process limited how quickly the audit team could deliver for customers. Solution: An external team was brought in to evaluate the company’s approach, optimize the existing model, and explore alternative model strategies to improve audit efficiency. The work included reassessing the model approach (including whether Q&A would outperform STS), benchmarking different model choices (including inference time), and refining training data to reduce training time. The team also identified workflow issues in the modeling process and implemented a two-phase iterative plan to improve model accuracy while updating ML infrastructure for faster iteration and deployment. Results: In ~2 months, the engagement increased the core model’s accuracy by 15% using advanced NLP and a limited dataset. The accuracy lift increased audit team throughput, raising auditor efficiency by 55%. A foundational ML infrastructure was also put in place to support rapid ingestion of new data, retraining, and deployment at scale.

Skills

Project Details

Litmos faced pressure to transform its learning platform under new ownership and leadership, while operating on a legacy codebase and managing hundreds of thousands of rigidly formatted courses. The team saw LLMs as both a disruption risk and an opportunity to leap ahead in learner outcomes. A major pain point was content discoverability across large libraries and formats, where keyword-based search forced learners to spend too much time navigating instead of learning. Litmos needed a more intuitive, conversational way for learners to find and use training content. Litmos partnered with Tribe AI to accelerate delivery of an AI-driven learner experience and internal content workflows. The teams built an AI Assistant to improve content discoverability and enable natural-language interactions, delivered in two phases: a proof of concept using 100 courses and two core workflows (“assign courses” and “see my courses that are due”), followed by an MVP spanning 300,000+ courses with multilingual support, secure access controls, and system/database integrations. They also embedded AI-powered content generation directly into the LMS, including a Module Generator that created multi-page learning modules and enhancements to the Content Authoring Tool with 12 AI-enabled components. The implementation handled complex ingestion and preparation across SCORM, audio/video transcription, PDFs, Word files, and native objects. Litmos launched the AI Assistant in May 2024 and launched AI-Powered Content Generation in January 2025. The company reported that launching paid AI features added millions in ARR and those features were adopted by 70% of new customers. AI-powered content generation reduced course creation time from hours to minutes and enabled instructional designers to draft modules in 3–4 minutes. Litmos also used early demos and staged rollout/testing to validate performance, security, and customer usage patterns before scaling broader go-to-market efforts.

Skills

Project Details

Bright faced a slow client onboarding process because its curriculum and immersive scenario development relied on a heavily manual, heuristics-based approach. Creating realistic customer situations for training required significant effort and limited how quickly new clients could be onboarded. Bright believed large language models could streamline and enhance training, but the team lacked a clear implementation path and domain expertise. Bright partnered with Tribe AI to design and prototype an AI-driven training experience powered by LLMs. Tribe AI staffed two NLP experts and worked through ideation, rapid prototyping, and scope crystallization aligned to Bright’s budget and timeline. The team built an AI chatbot that could adopt different customer personas and used OpenAI embedding, generative, and retrieval models to deliver dynamically generated customer simulations, a real-time CSR evaluation assistant, and a real-time CSR knowledge assistant. Bright later added a voice simulation layer to increase realism and interactivity. The resulting GenAI solution became a core part of Bright’s product demos and supported fundraising and sales efforts. Bright reported that customers who deployed the product said it increased ROI and decreased training costs. The collaboration also helped Bright move faster toward market entry and maintain a competitive edge. Bright planned additional work to improve latency and response quality as deployments expanded.

Skills

Project Details

Orchard Enterprise customers relied on data browsers with templates and filters to generate ad hoc lab reports. Those templates had a learning curve and required time and expertise to use efficiently. Orchard identified an opportunity to use generative AI to improve reporting efficiency and make the experience easier for end users. Orchard engaged Tribe AI to develop a proof of concept for a natural-language report generation interface (an “AI lab analyst”). The teams ran a discovery phase to align on scope, create a plan of action, and define success metrics, then moved into rapid prototyping. The prototype used AWS Bedrock with the Claude 2 large language model and was built with Python alongside Orchard’s existing environment. By the end of the 4-week POC, the prototype demonstrated a significant decrease in the time it took end users to create ad hoc reports. Orchard reported that the solution achieved all three success metrics defined for the prototype: faster report creation, a more intuitive user experience, and maintained reporting accuracy levels. The team also documented considerations for moving forward in healthcare, including patient perception, data security, and the required threshold for reporting accuracy.

Skills

Project Details

Consultants at a global consulting firm focused on financial and operational due diligence faced delays when reviewing unstructured documents like financial statements, contracts, and reports. Manually extracting key data was time-intensive and prone to errors. These bottlenecks limited how many documents could be reviewed per project. The firm wanted an AI approach that automated extraction and let consultants query documents directly. Tribe AI worked with the firm to implement an AI-powered document extraction and Q&A engine inside the firm’s proprietary due diligence platform. The solution combined OCR with large language models to convert scans and PDFs into machine-readable text and return answers to natural-language questions. It also extracted structured fields (e.g., financial figures, dates, vendor names, key terms) and mapped relationships across related contracts. The system provided source citations so consultants could trace answers back to the exact location in the original document. The implementation reduced document review time from 3–5 days to under 6 hours. It increased the number of documents reviewed per project by 60%. It eliminated 90% of the manual data entry previously required to extract key figures. The cited-answer workflow also reduced errors by improving auditability and traceability.

Skills

Project Details

Local news faced steep declines, with estimates saying 2 local papers closed every week and newsroom employment dropped 57% since 2004. Nota wanted to help publishers fight expanding “news deserts” by streamlining content creation and distribution. As Generative AI advanced quickly, the team needed to understand how to incorporate the technology into an actionable product roadmap. They also needed domain-specific guidance on SEO impacts, model choices, and differentiation in a crowded landscape. Nota partnered with Tribe to obtain immediate expertise and ongoing advisory support on Generative AI. Tribe assembled a two-person team: an LLM/SEO-focused senior data scientist and an NLP-focused data scientist/ML engineer to advise on models, interoperability, and architecture tradeoffs. The engagement included deep dives on GenAI’s impact on SEO, competitive landscape analysis, and ongoing roadmap guidance. This support helped Nota align technical decisions with product strategy and prepare for partner conversations. The work accelerated Nota’s ability to make informed model and architecture decisions and to shape a strategic AI product roadmap. Documentation and recommendations were used to quickly align the broader company and kickstart deeper technical exploration. The effort also supported a successful partnership meeting and contributed to Nota’s collaboration with Microsoft as part of the Microsoft Journalism Hub. Overall, Nota reported increased confidence in execution and decision-making as it moved into the next growth phase.

Skills

Project Details

Togal AI set out to modernize construction estimation (“takeoff”), a labor-intensive process that general contractors and subcontractors often duplicated across the same set of blueprints. The work was time consuming and typically required hours or days per set of plans, making bidding slower and more expensive. As a startup with non-technical founders, Togal also struggled to find scarce machine learning talent and initially spent time and money on an outside group that failed to deliver the promised progress. Togal partnered with Tribe to define a product approach that would be simple enough to drive adoption, focusing on a streamlined workflow (upload plans, click a few buttons, generate a report). Tribe assembled a cross-functional team spanning machine learning, data engineering, product, design, and full-stack engineering to build the system end-to-end. The team created a labeled dataset and a specialized labeling application, then built computer vision models to detect, label, and measure spaces and objects to AIA measurement standards. They implemented and scaled the platform on AWS using CI/CD and services including ECS/Fargate, ECR, RDS (PostgreSQL), EC2 GPU instances, Lambda, CloudFront, Route53, CloudWatch, and later custom inference on ECS behind load balancing and autoscaling. The resulting software automated estimating takeoffs in seconds—a process that previously took hours or days—reducing manual effort and speeding the bidding workflow. The AWS-based architecture allowed the system to scale with demand and operationally support bursts of usage. Tribe also helped Togal add dedicated technical leadership (a CTO) and provided ongoing advisory support beyond the initial build phase. Overall, Togal emerged with a faster, more streamlined product positioned to change how construction estimation is performed.

Skills

Project Details

Native Instruments needed to improve search and discovery across its large digital library of sounds, where creators used many different descriptors and search styles. The existing approach struggled to capture social context, which was important for interpreting pop-culture references and intent (e.g., a band name vs. literal words). The internal team specialized in ML and digital signal processing, not generative AI, and they already had a backlog of business-critical work. They therefore needed external expertise to fast-track a GenAI strategy without fully diverting internal resources. Native Instruments partnered with Tribe AI to run a four-week proof of concept focused on a human-like chat interface for sound search and discovery. The solution used Amazon Bedrock with Anthropic Claude 2, combined with Native Instruments’ proprietary indexed sound-file metadata and CLAP audio-text embeddings. Tribe AI built a full-stack cloud application (Django backend, React frontend) deployed on AWS with supporting services like ECR, EC2, S3, Route53, and ELB. The engagement emphasized discovery and scoping in the first half, followed by rapid prototyping in the second half to deliver a working prototype suitable for later A/B testing. The engagement produced a functional prototype chat interface designed to better handle social context in search queries. The new GenAI-assisted approach aimed to reduce the need for users to make multiple search attempts to get relevant results. Native Instruments shared the prototype and early outcomes internally and planned an upcoming A/B testing phase so leadership could evaluate it against the existing search experience. The project also helped the internal team learn GenAI concepts and validate that focusing on this deliverable was the right step for GenAI-centered innovation.

Skills

Project Details

Challenge: A leading private equity firm wanted to pressure-test a thesis that publicly available data could produce investable, predictive signals in a specific vertical. They needed an ML-driven toolkit that could score US geographies to improve how quickly and confidently they evaluated opportunities. The team also wanted flexibility on project cost and scope because they were not confident assessing technical talent. They required outputs that non-technical investors could use in real diligence and conversations with management teams. Solution: A three-person specialist team was assembled (technical product manager, ML engineer, and data scientist) and the work was structured into three sprints. The team gathered public datasets (e.g., census, social demographics, consumption patterns, infrastructure changes) and built a custom ML model to identify predictors and aggregate them into themes. After validating the model could extract signal from public data, they packaged the capability into an investor-friendly interface. They produced a US heat map, color-coded at the county level, to visualize the score and guide where analysts should focus. Results: The firm received an MVP within six weeks and used it to evaluate potential investments with increased speed and confidence. The toolkit improved how the team assessed a target company’s growth plan and how confident they were in key assumptions. It also helped the firm demonstrate value to prospective portfolio companies beyond capital by sharing public-data-driven insights. In some cases, those discussions contributed to more favorable deal terms and supported plans to expand the approach into additional datasets and verticals.

Skills

Project Details

Truebit Verify needed to determine how AI and machine learning could fit into its verified computing platform while maintaining transparency, security, and accountability in decentralized applications. The team faced open questions about whether ML training and inference could operate within Truebit’s verification protocol. They also needed clarity on how Truebit could contribute to AI verification as AI models became more outsourced and opaque. Overall, Truebit wanted actionable guidance to make the right product and business decisions in an AI-augmented Web3 landscape. Truebit engaged Tribe AI for expert advisory support in machine learning and generative AI. Tribe staffed Rahul Parundekar, an ML engineer with experience deploying models in security-sensitive environments and building MLOps pipelines. Together, they formed a strategic task force to identify key questions and knowledge gaps, supported by research and practical code examples. The work was organized into four weekly themes covering ML lifecycle fundamentals, identifying compute challenges and solutions, operationalizing models, and optimizing inference from deep neural networks to LLMs. Over four weeks, Truebit gained a deeper understanding of the ML landscape and how companies operationalized ML models in practice. The engagement helped the team evaluate options for providing compute to customers and expand their view of potential opportunities. Truebit reported that the guidance was actionable and enabled them to make informed decisions and correctly prioritize their roadmap. The team also validated its path forward with greater confidence and supporting data.

Skills

Project Details

Wingspan wanted to leverage machine learning as a competitive advantage to save its members time and money, but it did not need to hire a full-time AI team yet. The company needed its core product to become AI-enabled while staying aligned with platform-level priorities like risk and fraud detection. It also needed the right mix of financial services context, ML/NLP classification expertise, and Google Cloud Platform experience to execute quickly. Wingspan and a distributed Tribe-led team scoped a two-part engagement to be completed in 8 weeks. The team built and evaluated a proof-of-concept model to predict accounting categorization for transactions, classifying them as business vs. personal and then labeling business transactions as expense, income, or transfer with appropriate tax categories. They implemented the work on Google Cloud’s managed Vertex AI and collaborated closely with Wingspan stakeholders and technical leadership. Using learnings from the PoC sprints, they drafted a multi-year AI roadmap to prioritize ML use cases, data needs, and infrastructure investments. By the end of the 8-week project, Wingspan had a working transaction-classification PoC and an accompanying multi-year AI roadmap derived from the engagement. The roadmap clarified where Wingspan needed more consistent user-behavior data and better visibility into how users interacted with the platform. It also outlined incremental rollout of future ML use cases and identified data/infrastructure improvements intended to reduce bottlenecks and organizational load. Leadership reported the project delivered actionable direction for larger future AI investments.

Skills

Project Details

Boomi wanted to explore how generative AI could both support customer demand and improve how its services were delivered. The team needed to move from broad GenAI ideation to a focused proof-of-concept that solved a real customer problem. Customers often had to call the help desk or search FAQs and documentation to find precise answers. Boomi aimed to make answering highly specific product questions faster and more accurate using its existing corpus of manuals and support content. Boomi partnered with Tribe AI to run GenAI ideation and narrow the effort to a single POC. Tribe AI implemented an Amazon Bedrock-powered GenAI + LLM solution called the “Help Documentation Advisor.” The assistant accepted natural-language questions and responded using Boomi’s product manuals, help information, and community articles, including citations linking back to the source documentation. The system used a cloud-based full-stack architecture with an in-memory vector database (Chroma DB) and was deployed on AWS. Within 4 weeks, Boomi and Tribe AI developed an AI Advisor prototype. The prototype delivered responses that were faster and more precise than the traditional methods customers had used to find answers. Beyond the prototype itself, Boomi used the engagement to vet an AI-strategy development process alongside Tribe AI. Boomi also stated it gained a trusted partner it could recommend to its own customers for GenAI innovation and strategy work.

Skills

Project Details

Francisco Partners faced a fast-moving generative AI inflection point in early spring 2023 across a diverse set of portfolio businesses. Leadership needed a coordinated way to educate and inspire executives while also identifying practical GenAI opportunities and risks. The firm also wanted to accelerate adoption focused on product strategy and operational efficiencies. With 70+ portfolio companies, the challenge was scaling a consistent approach without relying on generic consultants. Francisco Partners formed an internal GenAI task force and partnered with an external team of hands-on GenAI practitioners. Together they created a series of tailor-made, industry-vertical ideation workshops designed to help portfolio leaders move from hype to actionable plans. The joint team then established a framework for enabling experimentation across the portfolio. This approach supported both strategic planning and execution through pilots alongside subject-matter experts. Less than 6 months after starting, the partnership delivered industry-vertical ideation workshops and established an execution plan. Over the next few months, the teams implemented over two dozen pilot projects with portfolio companies. Francisco Partners saw a significant increase in the number of portfolio companies undertaking AI projects and accelerated several initiatives that needed added GenAI expertise. The work also established an ongoing framework to guide future portfolio AI efforts as POCs moved toward production.

Skills

Project Details

Recursion, a clinical-stage TechBio company, faced performance bottlenecks in its BioHive supercomputing infrastructure supporting AI-driven drug discovery. The organization needed to improve GPU utilization while increasing overall training throughput. It also encountered challenges in distributed training and memory management that limited research velocity and raised compute costs. A team of 4 ML engineers was deployed to optimize BioHive for AI workloads. The work focused on GPU optimization to raise utilization and reduce waste. The implementation also addressed distributed training and memory management issues to improve end-to-end system performance. The engagement improved GPU utilization efficiency by 35%. It increased throughput by 10x, accelerating AI-driven research cycles. It also delivered $2.8M in annualized value through reduced compute costs and faster research.

Skills

Project Details

Follett needed to improve how K-12 students discovered reading materials in its Destiny Library Manager product. Students often struggled to find books that matched both their interests and reading levels. The existing experience did not sufficiently guide students to the right choices. Follett implemented an AI-powered chat feature within Destiny Library Manager to guide students toward appropriate reading materials. The solution generated personalized book recommendations based on a student’s interests and reading level. Experienced AI engineers were deployed and integrated with Follett’s team to deliver the recommendation capability. The implementation delivered an AI chat experience that helped guide K-12 students to relevant reading materials. Destiny Library Manager provided recommendations aligned to both interest and reading level. The engineering team integrated the solution into the existing product and delivered an updated discovery experience for students.

Skills

Project Details

Follett needed a better way for K-12 students to discover reading materials inside its Destiny Library Manager product. Students often struggled to find books that matched both their interests and reading levels. Follett also needed an approach that fit smoothly into an existing product and team workflow. An AI-powered chat feature was developed and implemented within Destiny Library Manager. The solution guided students through conversational prompts to surface personalized book recommendations based on interest and reading level. Experienced AI engineers were deployed to integrate with Follett’s team and deliver the recommendation capability within the product. The implementation delivered an AI chat experience that recommended books aligned to each student’s interests and reading level. The work integrated seamlessly with Destiny Library Manager and Follett’s existing team. The resulting recommendation system changed how students discovered books within the library platform.

Skills

Project Details

Recursion, a clinical-stage TechBio company, faced performance bottlenecks in its BioHive supercomputing infrastructure supporting AI-driven drug discovery. The organization needed to improve GPU utilization efficiency while increasing throughput for research workloads. It also sought to reduce compute costs and shorten research cycles. Recursion partnered with an external team to optimize its BioHive environment for AI workloads. A team of 4 elite ML engineers worked on GPU optimization, distributed training, and memory management. The implementation focused on improving how training jobs used hardware and scaled across the supercomputing infrastructure. The engagement delivered measurable performance and cost outcomes. GPU utilization efficiency improved by 35%, and throughput increased by 10x. The work generated $2.8M in annualized value through reduced compute costs and faster research cycles.

Skills

Project Details

Avalon Healthcare Solutions needed to automate prior authorization while maintaining strict accuracy in decisions. The process required extracting medical necessity criteria and cross-referencing it against clinical data. They also needed a way to handle complex cases that could not be cleanly automated. A GenAI solution was implemented to automate key prior authorization steps. The system extracted medical necessity criteria, cross-referenced clinical data, and used intelligent routing to send complex cases to the appropriate path. The delivery used a RAG architecture with fine-tuned LLMs. A team of 3 AI engineers built and delivered the solution. The implementation achieved production-ready prior authorization automation. It delivered 100% precision on automated approvals and reached an 83% recall rate. The automation improved efficiency by streamlining how criteria were extracted, data were checked, and edge cases were routed.

Skills

Project Details

Avalon Healthcare Solutions needed to automate prior authorization while maintaining high clinical accuracy. The existing process required extracting medical necessity criteria and cross-referencing it against clinical data. Complex cases also needed to be routed correctly rather than auto-approved. They sought a GenAI approach that could be production-ready. A GenAI solution was implemented to extract medical necessity criteria and cross-reference relevant clinical data. The implementation used a RAG architecture with fine-tuned LLMs to support accurate decisioning. An intelligent routing system was built to escalate complex cases appropriately. The work was delivered by a focused engineering team. The solution achieved production-ready automation for prior authorization workflows. It delivered 100% precision on automated approvals. It also reached an 83% recall rate. The automation improved efficiency by reducing manual effort in routine approvals while routing exceptions for review.

Skills

Project Details