Enterprise Knowledge Synthesizer & Decision Engine

The SwissHacks 2025 challenge: companies sit on unstructured data spread across silos — PDF, Word, Excel, PPTX, Markdown, plain text — and need to turn it into correlated business decisions, not just search results. The real difficulty was never the AI call. It was the two things around it: preparing that heterogeneous mess efficiently before handing it to a model, and stopping API costs from exploding once you do.

What we built#

We called it Finetic, our entry for the Arch/Re challenge. An end-to-end prototype for intelligent data aggregation and visualisation, built around three ideas.

1. Algorithmic data synthesis (before the AI)#

Rather than throwing raw documents at a model, we designed algorithms to structure and filter massive, heterogeneous datasets first, so the AI receives precise, context-relevant information. Most of the engineering — and most of the value — lives in this pre-processing layer, not in the prompt.

2. Human-in-the-loop that adapts in real time#

The part I'm proudest of: users give the AI feedback on proposed decision points directly from the frontend, and the system adjusts its reasoning strategy and how it correlates data in real time. It is a control surface over the model's thinking, not a one-shot answer you either accept or discard.

3. Cost as a first-class constraint#

We built a backend strategy specifically to minimise the number of tokens shipped to the LLM APIs — because a decision engine nobody can afford to run is not a product. Cutting token transfer was treated as a core feature, the same way the Sentinel work a year later made cost structural.

The result#

A minimalist, deliberately simple frontend that translates complex analysis into direct, actionable insights — the complexity pushed into the pipeline so the user only ever sees the decision.

Stack: Python · FastAPI · OpenAI API · Next.js · Vercel.