AI Engineer
Video Knowledge System
Video knowledge extraction with a real eval harness. Node.js + TypeScript on the Anthropic API.
A Node.js and TypeScript pipeline on the Anthropic API for video knowledge extraction. JSON Schema defines the knowledge atom shape with evidence validation on every persisted record. Real eval harness with eval-retrieval-multi, schema-conformance QA, and baseline snapshot scripts for regression detection.
The problem
Extracting knowledge from video into a structured, queryable form is a common AI demo, but most implementations cannot tell you whether the extraction is actually correct. Without a real eval harness, regressions land silently. A schema that lacks evidence validation lets the model claim facts that have no provenance back to the source.
What we built
The pipeline runs in Node.js and TypeScript on the Anthropic API. The knowledge atom shape is defined in JSON Schema with evidence validation on every persisted record. The eval harness covers retrieval quality across configurations (eval-retrieval-multi), schema-conformance QA on extracted atoms, and baseline snapshot scripts that catch regressions when prompts or models change.
What was delivered
- Node.js + TypeScript extraction pipeline on the Anthropic API
- JSON Schema definition of knowledge atoms with evidence validation
- Eval harness: retrieval quality, schema conformance, baseline snapshots
- Regression detection that fires when prompts or models change
Outcomes
- Extracted atoms are auditable; every fact ties back to source evidence
- Regressions are caught at the eval harness, not in production
- The testing discipline that separates a demo from a system a team can ship
- Reusable pattern for retrieval-and-extraction pipelines on other source media
Services: TypeScript, Node.js, Anthropic API, JSON Schema, Eval harness