The problem
Getting a business request to shipped code meant a chain of manual handoffs: write the spec, decompose it, implement, write tests, write documentation. Nothing owned the whole chain, so the honest answer to "where is this task" was a conversation rather than a query.
What I built
I initiated the platform — I wrote the repository's root commit — and have led it as maintainer while it grew to roughly 8 contributors and ~116K lines across a FastAPI backend and a Next.js 15 / React 19 frontend, merging 94 feature branches to master along the way.
The components I built and own end to end:
- The dual-engine LLM provider abstraction — one interface over two incompatible coding-agent runtimes (a forked CLI binary and an in-process SDK), swappable by a single Parameter Store value.
- The subprocess layer that makes forking a 600-second child safe.
- The Atlassian OAuth 2.0 + JWT session layer — the auth foundation every integration on the platform authenticates against.
- The natural-language → security-validated read-only API workflow and its self-healing validate→repair loop, with semantic and syntax back-edges.
- The SSE streaming path that surfaces multi-minute agent runs to the browser.
How it works
16 workflow types are modelled as a persisted phase machine over a 46-table MySQL schema — deliberately without a LangGraph checkpointer, so one store answers "where is this task" instead of two. Each phase is a LangGraph graph; transitions are gated on human approval. Long agent runs execute as forked subprocesses and stream progress to the UI over SSE, with phase events published to Kafka (KRaft) for downstream consumers.
Evidence
Root commit and maintainership · 94 feature branches merged · 16 workflow types · 46 tables (exact) · ~116,486 LOC platform size · Terraform deployment of 44 resources across 26 AWS types.
What I owned
I initiated this platform and lead it. I designed the persisted phase machine and the workflow model, and personally built the engine core and its self-healing validate→repair loop, the subprocess manager, the Atlassian OAuth 2.0 + JWT auth layer, the dual-engine provider abstraction, and the SSE streaming path. I operated the AWS deployment behind it — Terraform across 44 resources and 26 AWS types — and reviewed the work that came in from the rest of the team. The ~116,486 LOC figure is the platform's size, not a personal line count.
Stack
FastAPI · LangGraph · Kafka (KRaft) · MySQL · Next.js 15 · React 19 · TypeScript · Terraform · AWS (EC2, ASG, ALB, ACM, RDS, ECR, SSM)