Human API
A real-time platform that turns multi-user team chat into a versioned, executable work contract.
Technology: Gemini, REST, WebSocket, JWT, PostgreSQL, MongoDB, pytest, Vitest.
Overview
Human API turns multi-user team chat into a versioned, executable work contract. It uses an LLM-as-parser pipeline: Gemini extracts intent but never owns truth. Deterministic engines decide whether that intent is ready to leave the conversation, so contradictions cannot silently overwrite agreed values.
Problem
Team chat does not, by itself, produce a work agreement that can be versioned, checked, and handed to a delivery system.
Challenge
A model can extract intent from conversation, but it cannot be the source of business truth. Extraction and the rules that accept an agreement have to stay separate.
Architecture
- Gemini extracts structured intent from the conversation.
- Deterministic merge, readiness scoring, conflict detection, and verification engines decide the agreement state, so contradictions cannot silently overwrite agreed values.
- A status machine tracks each agreement from draft to complete.
- REST and WebSocket APIs expose that state. Authentication uses JWT.
- Dual storage: PostgreSQL and MongoDB persist the product data.
- pytest and Vitest test the verification logic and the interface.
- Work packages, and optional Jira tickets, are emitted only when the agreement is complete and conflict-free.
Contribution
Designed the LLM-as-parser pipeline and built the real-time REST and WebSocket collaboration, the status machine, and the APIs that store and deliver a conflict-free result.
Outcomes
- Work agreements are structured, versioned, and verified before they leave the conversation.
- Jira tickets and work packages are created only after the agreement is complete and conflict-free.