USense

My role: one of a three-person team. I improved RAG retrieval, and I designed the architecture of the NL2SQL feature and built it.

Part 1: Making document answers reliable

USense began as a simple RAG chatbot running on any local model Ollama supports, for teams whose documents couldn't leave their own machines. The model was swappable; what mattered was the harness around it, the context and tools that let a small local model give good answers. Its only controls were top-K, top-P and temperature. It worked on single, simple documents, and broke when a question spanned several files, or when the answer was spread across chunks as tables, diagrams or numbers.

What I added:

1. HyDE: write a hypothetical answer first and search with it, so questions match the language of the documents
2. Query embeddings tuned for retrieval
3. Variable top-K: simple questions pull a few chunks, broad ones pull more
4. Tuned chunk size and overlap, so tables and numbers stop getting split mid-thought
5. Parent-child chunking: search on small precise chunks, answer with their larger parent sections
6. Multi-document search, so one question can draw on several files

Result: retrieval hit rate from 58% to 94%.

How it was measured: a custom evaluation on free-to-use documents: published research papers, technical textbooks and novels. Each document was chunked at several sizes and a question written for every chunk. The documents were then uploaded through the app as a normal user, every question was run, and a hit meant the answer chunk appeared in the top-K retrieved chunks.

Part 2: Answering from the database

Who it was for: business leaders in finance, retail, marketing and sales who check the data before they decide. Instead of sending a question to an analyst and waiting on their turnaround, they ask USense directly. When an answer deserves a permanent dashboard, they hand the analyst the USense result as the spec, so the dashboard starts from a clear example rather than a vague request.

How it works: a user asks a business question in plain English. USense turns it into SQL, runs it, and replies with a summary, the SQL and the rows it used, so people can check the work. Answers export to CSV, PDF, Excel, Google Docs, Power BI and Tableau.

How it got to 84%, one problem at a time:

Models: local models for the individual-user version; Mistral Large, hosted in the cloud, for the enterprise version. Both reached similar accuracy: local models on simpler schemas, Mistral Large on complex enterprise schemas.

Safety and cost: role-based access, a SQL parser that blocks every non-read statement before it runs, BigQuery partitioning and clustering guidelines, per-query cost limits and audit logging.

What came next: USense was retired as frontier models caught up. Its ideas live on in PlainQuery.