DeepSource
Staff Software Engineer
Oct 2021 — Present
- Led a five-engineer AI team — set the technical direction, make the calls on architecture, and mentor the group, while still spending most of my time hands-on in the codebase.
- Built our multi-agent code-review and remediation pipelines on Pydantic AI, with code-extraction tooling and static-analysis-driven context shaping the prompts — 84.51% F1 on vulnerability detection against a public benchmark.
- Shipped a secrets-detection engine that fine-tunes small LLMs for the task — high precision (92.78% F1) without the cost profile of a frontier model.
- Designed long-term memory for our agents on top of embedding-based retrieval, so each review draws on the relevant history of that codebase instead of starting cold every time.
- Built the evaluation and observability layer underneath all of this: agent-level performance metrics, traces across multi-step runs, plus latency, failure-mode, and token-spend monitoring in production.
- Wrote DeepSource's JavaScript and Swift static-analysis engines from the ground up, with extensible APIs for rules and autofixes. The work cut analyzer failures by 63% and grew usage by 118%, now serving 10K+ customer codebases across 200K+ monthly runs.