Hands-on build-along

Build a Contract-Intelligence Pipeline

Build a working contract-intelligence pipeline end to end. Ingest sample contracts, run OCR + open-source LLM extraction, persist to Postgres against a realistic contract data model, build an evaluation harness, expose results via API. ~3 hours start to finish; ~6 chapters of working code at every step.

What you'll build

A minimal but real contract-intelligence pipeline. Drop a PDF in a folder; an ingestion job picks it up, runs OCR, extracts structured fields with an LLM, persists to Postgres with the contract data model, and exposes results via a FastAPI endpoint. An eval harness scores accuracy against a small gold set. Working code; ready to extend.

Prerequisites

Python 3.11+, Docker (for Postgres), an LLM API key (Anthropic Claude or OpenAI — Claude shown by default), familiarity with FastAPI or willingness to read along, and a few sample contracts (5–10 PDFs work; public templates are fine for a first pass).

Setup

00Start Here What you'll build, prerequisites, the shape of the project, how to use this guide. 01Project Setup Repo structure, virtualenv, dependencies (FastAPI, SQLAlchemy, anthropic SDK, pytest), Docker Compose for Postgres. Hands-on · 20 min

Build steps

02Ingestion: Folder Watcher to Postgres Watch a folder for new PDFs. Compute a stable document_id (content hash). Persist file metadata in documents. Handle re-ingest idempotently. Hands-on · 30 min 03Extraction: OCR + LLM Use Tesseract for OCR on scanned PDFs; route native PDFs through pdfplumber. Then call Claude with a structured-output prompt to extract parties, effective date, renewal terms. Confidence scoring. Hands-on · 45 min 04Persistence: The Contract Data Model SQLAlchemy models for documents, agreements, obligations, suppliers. Upsert by (document_id, model_version) for idempotency. Provenance columns. Hands-on · 30 min 05Evaluation Harness A tiny gold-set repository, a CLI to run extraction over the gold set and score per-field accuracy, a calibration check, a Markdown report. Run before every prompt or model change. Hands-on · 30 min 06API Exposure FastAPI endpoints: list agreements, fetch an agreement with its obligations, search by supplier. Pagination, filters, basic auth. The API your customer's BI / app would call. Hands-on · 25 min

Extensions

07Extensions & Next Steps Where to grow this — layout-aware models, HITL workflow, Snowflake export, dbt models on top, multi-tenant, observability. Pointers into the deep dive and the playbook.