DocIntel AI — Multimodal Document Intelligence Platform
A multimodal document intelligence API that extracts structured data from invoices and receipts using vision LLMs, then validates every result against strict Pydantic schemas and deterministic business rules before it's trusted.
Vision LLMs are good at reading documents but not reliable enough to trust blindly — they hallucinate totals, misread dates, and produce inconsistent field names from one run to the next. Most document-extraction tools either accept whatever the model returns or require a human to review every single result, neither of which scales.
DocIntel treats the LLM as one fallible stage inside a deterministic pipeline, not the source of truth. Every uploaded PDF or image is preprocessed and page-split, then sent to Gemini as the primary vision model with GPT-4o as an automatic fallback. The raw response is parsed against strict Pydantic v2 schemas, checked against deterministic business rules — line-item totals, date validity, currency consistency — and scored for confidence, auto-accepting high-confidence extractions and flagging the rest for human review. The pipeline is built as single-responsibility FastAPI services with thin routes, backed by SQLAlchemy/PostgreSQL for persistence and a 68-test suite with mocked LLM providers.
The result is a document-extraction API that fails safely instead of failing silently: uncertain fields get flagged instead of shipped, and a bounded retry/fallback chain between two vision providers keeps the pipeline running even when one model is degraded or unavailable. A demo UI and evaluation harness make it straightforward to validate accuracy against a synthetic dataset before trusting the pipeline on real documents.
Figures are from our own build and R&D benchmarks — not independently audited client results.
