I build across the stack: backend APIs, full-stack web apps, AI features, and MCP tooling. I lean toward projects that get tested against real conditions, not just demoed once and shipped.
- 🧪 I test things by actually breaking them: race conditions under real concurrent load, clean-checkout runs, chasing down query-planner behavior instead of trusting the first benchmark number.
- 🛠️ Backend: FastAPI/Flask APIs, PostgreSQL. AI: LLM-powered features, Anthropic/OpenAI SDKs. MCP: building servers and tools that let models act on real systems. Full-stack: React + TypeScript. Infra: Docker, AWS (EKS, multi-tier deployments). Also comfortable in C++.
- 🎓 Software engineering student, currently exploring internship / working student opportunities.
- Found and fixed two SQLite concurrency bugs in DBOS (merged): workflows sharing a concurrency budget could overspend it, and a transaction step could write-skew past a cap, both from reads running outside the write lock. Proven with deterministic concurrency tests, not timing-dependent ones.
- Closed a silent migration race in DBOS (merged):
DBOS.migrate()would log a warning and run anyway if it couldn't get its advisory lock in time, letting two migrations run concurrently. Scoped the fix with the maintainer first, then shipped it with a test that holds the lock from a second connection to prove the failure. - Licensed DevPulse into the XERJ Corpus Hub: an open reference library AI coding agents search before writing code, indexing a real commit (per-org rate limiting, bounded Claude request timeout/retries) as a worked example.
- Fixed a dtype-validation gap in kornia (merged): two geometry conversion functions let integer/bool tensors through to a bare, unhelpful
RuntimeErrorinstead of a clear error matching their sibling functions. - Fixed a mistagged register in Home Assistant's solax-modbus (merged) integration: three X1 Boost G4/Smart G2 entities were tagged for the wrong hardware generation (GEN3 instead of GEN4), so Manufacturer/Model/Language never read correctly on that hardware.
- Fixed dill's (merged)
check()failing on large objects: it built a subprocess command by embedding the pickled object in a shell-split string, blowing past the OS argv length limit. Switched to passing argv as a list and piping the payload over stdin instead. - Fixed wrong AND/OR wording in cron-descriptor (merged) for combined day-of-month/day-of-week expressions: traced the actual behavior to Vixie cron's source rather than just the POSIX spec, matching its documented-bizarre bug where a day-of-month field starting with
*(e.g.*/2) gets ANDed with day-of-week instead of ORed. Shipped with regression tests covering both the fixed case and the preserved buggy-but-correct case.



