Document AI · 2025–2026

DocuIntel

A document intelligence platform that turns messy, mixed-format documents into validated, structured data. Users define the fields they need, pick any of six LLM providers, and keep their own keys encrypted end to end.

Role: Own project, from design to deployment
Stack: Next.js · FastAPI · LangChain · Supabase

01 ProblemManual data entry and brittle OCR pipelines, locked into fixed schemas and per-document fees.

02 ApproachUser-defined fields compiled into schemas at runtime, run on any of six LLM providers.

03 TechnologyNext.js 16, FastAPI, LangChain structured output, Supabase with row-level security.

04 OutcomeZero hardcoded schemas, six providers, and bring-your-own keys encrypted with AES-256-GCM.

01 — Problem

Teams lose hours to manual data entry and OCR pipelines that break on every new layout. Commercial tools lock them into fixed schemas, charge per document, hold their API keys centrally, and struggle when PDFs, scans, CSVs and plain text arrive together under load.

02 — Approach

01Ingest any format without blockingPDFs, CSVs, text and images stream in through secure multipart upload. Files are sanitised against path traversal, processed in isolated temporary storage, and parsed off the event loop so large uploads never stall the API.

02Schemas built at runtime, not in codeThe fields a user defines are compiled into a Pydantic model on the fly. LangChain's structured output then enforces that schema on OpenAI, Claude, Gemini, Groq, Mistral or Cohere, so switching provider never means rewriting extraction rules.

03Zero-trust keys and a human in the loopUsers bring their own AI keys, which are never stored unencrypted: AES-256-GCM with PBKDF2 key stretching and PostgreSQL row-level security. A review dashboard lets people correct results inline, save templates and export the data.

03 — Technology

Frontend: Next.js 16, React 19, TypeScript, Tailwind CSS v4, shadcn/ui
Backend: FastAPI, Python, Pydantic v2, pypdf, pandas
AI: LangChain structured output across OpenAI, Claude, Gemini, Groq, Mistral and Cohere
Data & auth: Supabase PostgreSQL with row-level security, SSR auth with HttpOnly cookies
Security: AES-256-GCM, PBKDF2, Cloudflare Turnstile, rate limiting
Delivery: Docker multi-stage builds, native iOS app (SwiftUI)

04 — Outcome2025–2026

LLM providers on one runtime schema engine

PBKDF2 rounds protecting every stored key

hardcoded extraction schemas in the codebase

Next project · 03ScopeAISix AI agents turn a one-paragraph product idea into a full technical blueprint in under a minute.