CS + Business @ Carnegie Mellon '27 · RL environments, agent evals, and agent infrastructure
- pipeline-reliability: a generator of 282 validated Harbor tasks in 56 problem families where agents fix reliability bugs in LLM data-pipeline code. Every task passes an automated QA gate (reference 1, unfixed 0, wrong/decoy/tamper fixes 0); the grader runs unprivileged in a separate container and writes the reward last.
- claims-vs-reality: a pre-registered study comparing coding agents' final reports with an exact system-call record of what they ran. Rubric judges caught at most 1 of 3 misreport mechanisms; the exact record caught all of them.
- fin-model-edits · filing-update-evals: exact-graded agent evals on real SEC filings and credit agreements, recalculated in LibreOffice with perturbed inputs; no LLM judge.
- Merged reliability fixes in agent infrastructure: Cua, Daytona, Osmosis, Composio (co-author).
- Airbnb, AI Infrastructure intern: built a conversational agent for an internal prompt-engineering platform, with RAG, a confirm-before-apply step, OpenTelemetry tracing across the inference pipeline and an eval harness. Full-time return offer.
- Parameter Golf: 16MB language model trained on 8×H100s with 60+ tracked ablations; briefly topped the leaderboard. The score was flagged in a shared upstream train/val overlap issue, so I treat it as optimistic.
- Also: Relay, a long-horizon agent with crash-safe MongoDB checkpoints (MongoDB NYC hackathon).
Go · Python · TypeScript · C/C++ · Docker · Harbor · PyTorch/CUDA · vLLM · gRPC · PostgreSQL
Open to full-time roles starting January 2027 · Website · LinkedIn · tgudise@andrew.cmu.edu