Generating high-quality insights from this market intelligence platform's data traditionally required Senior Analysts with significant experience, who were limited to primarily responding to client-initiated questions. Important market signals, such as internet trends and consumer sentiment, often preceded measurable sales shifts, but there was no scalable way to detect or validate them. The company sought to proactively identify and validate trends their clients hadn't yet considered using proprietary point-of-sale data. Fractional AI built a modular research agent system designed to mimic the workflow of an experienced Analyst, operating in two stages: topic generation and structured evaluation. Agents generated candidate 'areas of concern' by scanning public web data for emerging trends. In the second stage, a multi-agent debate analyzed each topic, with two AI agents arguing for and against the trend's importance using sales data and external web evidence, followed by a third agent simulating a brand executive to render a final decision. This trend report agent set the foundation to shift from reactive analytics to forward-looking and insight-driven consulting. Reports were generated in under 35 minutes, enabling rapid delivery of insights to clients, and each report cost approximately $3 to produce. The system enabled proactive detection of early trends backed by internal data, laying the groundwork for future productization and scalability, with approximately 80% consistency in report verdicts.

Skills

Project Details

Sincera, a company decoding digital advertising data, faced a major challenge with millions of messy, inconsistent, and unstructured product records collected monthly from various internet sources. This data was nearly impossible to use without extensive and time-consuming standardization. The goal was to map these records to Shopify’s product taxonomy, which included 10,000 categories and up to 7 levels of nested hierarchy, a feat that would be cost-prohibitive and require thousands of human hours without advanced AI. To address this, Sincera partnered with Fractional AI to build an AI categorization system using a multi-step LLM pipeline. This system evaluated each record to classify it as a brand, segment, or uncategorizable, then enriched sparse records and discerned the best matching category. The solution outputted each record to its corresponding Shopify category with a level of confidence in real-time, effectively transforming the unstructured data stream into a valuable asset. Additionally, the collaboration significantly enhanced the Sincera team's confidence and skills in AI engineering, particularly in building robust LLM evaluations. As a result, Sincera's monthly stream of messy data was successfully transformed into a valuable, usable data asset. The AI system consistently achieved categorization accuracy above 85% for millions of records. Each record was processed and categorized in real-time, providing immediate data utility and unlocking insights that were previously inaccessible due to the data's complexity and volume.

Project Details

Superintelligent, an AI learning platform offering hundreds of daily updated tutorials, faced a challenge in helping its learners efficiently navigate the vast content. Users struggled to quickly find relevant AI tools for their specific needs through traditional search methods. This was particularly difficult when they didn't know the names of specific tools or if certain solutions even existed, creating a significant barrier to an efficient learning experience. Fractional AI partnered with Superintelligent to address this by building an AI chatbot leveraging Retrieval Augmented Generation (RAG). This chatbot was designed to offer members specific, personalized tool recommendations based on their queries. The solution was engineered for perpetual refreshment, allowing the Superintelligent team to seamlessly add new tools to its underlying database, with updates automatically syncing to the chatbot to ensure current recommendations. The implementation of the AI chatbot successfully provided hundreds of personalized recommendations to learners. This significantly reduced the time users previously spent searching for relevant AI tools, allowing them to dedicate more effort to actual learning. The solution effectively fostered a more personalized and efficient user experience on the Superintelligent platform.

Project Details

A PE-backed e-commerce company faced an expensive document processing task that relied on a business process outsourcing (BPO) firm. Their workflow involved extensive manual review of lists, each containing nuanced information that required structured mapping and validation. This annual, large-scale process incurred substantial BPO costs, hindering efficiency and product improvements. Fractional AI designed and deployed an end-to-end automated system powered by large language models. The new process extracted key information from raw documents, structured data into standardized formats, and mapped items with associated quantities and annotations. It also evaluated items to determine human QA requirements and automatically incorporated feedback for self-learning. The system architecture was built for scale and reliability, featuring multiple levels of retries and timeouts to ensure continuous operation. The new genAI-powered system became 84% cheaper to run, with costs expected to decline further. List processing time dramatically dropped from over 24 hours to just 53 seconds. QA workload was immediately slashed, allowing internal QA time to be reallocated to higher-priority projects. The model delivered more accurate results than legacy workflows, with QA costs projected to approach $0 through incorporated feedback.

Project Details

Beghou faced challenges scaling its Beghou Arc platform due to the extensive manual effort and specialized engineering expertise required for custom client configurations. Each instance involved configuring hundreds of interrelated metadata tables and over 1,500 columns, making even minor updates complex. This manual process limited scalability and slowed project delivery times, impacting the ability to onboard new engineers efficiently while maintaining high standards. Beghou partnered with Fractional AI to develop the Beghou Arc AI Copilot, an intelligent system designed to automate complex configuration workflows. This AI copilot translates natural-language requests from engineers into validated SQL proposals without executing changes automatically. Integrated with the platform's UI, it inspects current configurations and generates scripts, ensuring expert oversight while automating repetitive and error-prone aspects of Arc configuration. The implementation of the AI Copilot dramatically improved efficiency, enabling configuration tasks to be completed up to 10 times faster, reducing project delivery timelines from days to minutes. This automation allowed senior developers to focus on innovation and advanced problem-solving. The system also ensured consistency in SQL proposals, accelerated new engineer onboarding, and established a scalable foundation for future AI-assisted configurations across Beghou’s portfolio.

Project Details

Superintelligent sought to scale its agent readiness audits, a crucial process for identifying AI opportunities within large organizations. Traditional methods, relying on consultant-led, high-touch interviews, proved prohibitively expensive and difficult to scale. Static surveys also lacked the adaptability required for deep qualitative research and trust-based conversations. The company aimed to develop an AI voice agent capable of conducting autonomous, human-like interviews at scale, with 24/7 availability and the ability to run hundreds simultaneously. Fractional AI developed a production-ready voice agent, built on a modular architecture and a customized real-time LLM orchestration system. This agent enabled Superintelligent to conduct adaptive, on-demand interviews at scale, generating deep transcript-level insights. Key features included smart follow-up questions, the ability to conduct simultaneous interviews with hundreds of employees, and real-time transcript filtering and aggregation. The voice agent conducted thousands of interviews in its first month of production. At one Fortune 100 client, it interviewed 150 employees within two weeks for a total cost of $500, delivering a 10x better experience than previous methods. This enabled Superintelligent to scale rapidly and is prepared to support millions of interviews annually, unlocking information gathering at an unprecedented, low-cost scale.

Project Details

Change.org faced the challenge of moderating thousands of daily campaigns to meet community guidelines, requiring a system to manage content violations like hate speech and misleading information. Their existing no-code solution, built with Google Sheets, Zapier, and GPT, automatically flagged about half of content violations but struggled with a 50% false positive rate for petitions sent for manual review. This meant significant human time was spent reviewing harmless content, and many dangerous petitions still fell through the cracks. They aimed to move beyond spreadsheets to a more robust, integrated system that caught more violations and reduced false positives. Change.org partnered with Fractional AI to develop a more robust and scalable AI-powered content moderation system. Through over 100 experiments, Fractional AI created an AI system designed to navigate the nuances of content moderation. This new system was integrated directly into Change.org's tech stack, replacing the previous spreadsheet-based workflow with a streamlined REST API workflow. The solution involved fine-tuning GPT 3.5, using the OpenAI API moderation endpoint, and incorporating structured output with chain-of-thought prompting. The implemented AI system significantly improved content moderation, detecting 77% of content violations. It also dramatically reduced the false positive rate, cutting it in half (from 1 in 2 to 1 in 4 petitions flagged for review). This reduction in false positives meant human reviewers spent their time more efficiently. The system replaced a complex spreadsheet workflow with a robust REST API, seamlessly integrating into Change.org's environment, all while keeping daily operational costs around $30.

Project Details

Airbyte faced the challenge of scaling its data connectivity by building thousands of new API integrations, or "connectors." The manual process of building these connectors was tedious and complex, requiring developers to navigate diverse API documentation and configure numerous fields. This not only consumed significant time but also diverted valuable technical talent from higher-value work. Airbyte partnered with Fractional AI to develop an AI-Powered Connector Builder. This solution enabled the AI to crawl API documentation, automatically pre-populate connector fields, and present a list of available data streams. Users could then select desired streams, and the AI would configure the necessary fields in Airbyte's UI, allowing for final review and edits. The AI-Powered Connector Builder drastically reduced the time to build a connector from hours to just a few minutes. This significantly lowered the barrier to creating new connections, enabling Airbyte to expand data connectivity across a greater number of sources. The company observed a marked increase in the volume of new connectors after the AI Assist release.

Project Details

Zapier utilized a "spec gen system" powered by large language models to automate the creation and maintenance of OpenAPI specifications for over 7,000 applications. While the system was effective, Zapier lacked clear visibility into its performance and accuracy. They needed to establish robust methods to measure the system's baseline accuracy, quantify the impact of engineering changes, and strategically guide future improvements. This lack of precise measurement hindered their ability to optimize the LLM pipeline efficiently. Fractional AI collaborated with Zapier to implement a two-part solution. First, they defined robust LLM evaluations by developing a data-driven framework to precisely measure the accuracy of each subcomponent within the spec gen system. This framework assessed aspects like the incidence of hallucinated paths and the accuracy of endpoint detection. Second, with a clear measurement standard established, iterative experiments were conducted. These experiments systematically tested various changes in model selection, prompting techniques, and pipeline ordering to enhance the spec gen system's performance. The partnership resulted in dramatic improvements across key metrics for Zapier's spec gen system. The incidence of hallucinated paths was significantly reduced from 26% to less than 1% following the improvements. Additionally, the accuracy of automatically detecting field types nearly doubled. These enhancements led to a more reliable spec gen system, enabling more integrations and substantial time savings for Zapier's engineers. Cost-saving interventions also included prompt caching, which saved approximately 25% per run.

Project Details

Cando Rail & Terminals aimed to multiply their annual risk assessments sixfold while ensuring human experts retained full control over safety decisions. The existing manual process, however, presented significant scalability challenges. Each assessment typically required up to 16 hours of SME meetings and weeks of revisions, often involving tedious PDF cross-referencing. This method was inefficient and offered limited opportunities to learn from previous assessments. Fractional AI partnered with Cando to develop "Peter the Safety Agent," an AI-powered risk assessment system. This solution automated documentation and streamlined collaboration. The system's prework automation feature produced initial drafts of risk assessments in mere seconds. Additionally, an in-meeting voice assistant transcribed and structured expert discussions in real-time, generating updated drafts as conversations progressed. The implementation of Peter the Safety Agent resulted in significant time savings, reducing assessment time by over 20 hours per assessment. Reports were drafted in seconds for less than $0.05 each. This innovative system enabled Cando to expand its risk assessment program sixfold in just 17% of the time previously required. Crucially, human experts maintained full control over every safety decision throughout the process.

Project Details