Our product: DocumentInsight
Most document search returns a file and leaves the reading to you. DocumentInsight takes a PDF — digital or scanned — extracts its text, files it into a searchable knowledge base, and answers questions about it with page-level citations. Two cooperating agents do the work, and a deterministic check rejects any citation the system did not actually retrieve.
Open the app ↗Real documents are not clean. A contract arrives as a scan, a report hides its numbers in tables, and a policy is revised three times with every version still in circulation. A pipeline that runs one fixed extractor fails silently on the first two and contradicts itself on the third.
An answer is only useful if you can check it. Every claim had to point at the document and page it came from — and when a new document disagrees with an old one, the system had to notice rather than quietly keep both.
Two LangGraph agents hand off inside one ingestion graph. The first chooses its own extraction strategy — PyMuPDF first, pdfplumber for tables or scrambled text, OCR for scans — judging a quality report after each attempt and escalating until the text is usable. The second chunks the text page by page, compares it with what the knowledge base already holds, and resolves conflicts explicitly before indexing into Qdrant. Ingestion runs in Celery workers, so a large upload never blocks the interface.
Live since 12 September 2026, with 20 endpoints and 233 automated tests — 126 on the backend, 107 on the frontend.
Maria Vlahova-Takova designed the agent architecture and wrote the ingestion, chat and memory agents and the first API. Stoyan Takov built the interface, took the product to production, and has written most of the code since.
Stack: FastAPI · LangGraph · Qdrant · Celery · Redis · PostgreSQL · OCRmyPDF · React 19