Merged into PyTorch and AutoGPT.
I hunt correctness bugs in code other people depend on.
One PyTorch bug, taken apart
On Apple GPUs, torch.nextafter returned bfloat16 inputs unchanged. I rewrote it as a step on the raw bits.
Ask for the next number up. The old code hands back the same number; the fix moves one step. Press Step to see it.
Before my fix
1
0x3F80
After my fix
1
0x3F80
Step 0
Fixed it. Built it.
AutoGPT
Blank schedule names crashed the backend.
Mergedcodebase-memory-mcp
Cut 2,056 false call edges in a benchmark.
MergedAtlas
Idea to architecture, each stage approved.
22 of 22 evalsTraffiTwin AI
Rebuilds dead traffic-sensor speeds, 6.0% error against a 28.1% baseline.
Live demoPayPact
Policy gate for AI agent payments.
217 testshealth-policy-agent
Policy answers checked against clauses.
265+ tests
What I take on
Open to remote roles in backend and ML systems. Send me the bug nobody can explain.
Correctness fixes
Trace a wrong result to its cause, then fix it with a small change and a test. Merged into PyTorch and AutoGPT.
Backend services
Python and FastAPI services on PostgreSQL, tested as I build. PayPact has 217 backend tests.
AI and ML systems
Forecasting models and LLM agents, measured rather than assumed. TraffiTwin reaches 6.0% error on METR‑LA against 28.1% for the historical average.
Evaluation suites
Benchmarks checked against a stored baseline, so a regression fails loudly. Atlas passes 22 of 22 eval tasks, run with a local model.
Database engineering
Audit logging, backup verification and health monitoring on PostgreSQL and SQL Server, built as a database intern at DCM Infotech.
Honest status
I say what is done, what is still in progress, and what did not work.
Tools I reach for
- Languages
- PythonC++JavaScriptSQL
- Backend
- FastAPIREST APIsNext.js
- Databases
- PostgreSQLSQL ServerOraclePL/pgSQL
- ML and data
- LightGBMscikit-learnMLflow
- Tooling
- DockerGitGitHub Actions
