Software engineer, CS new grad (December 2026), Charlotte NC. I work on AI agent reliability, applied ML, and benchmark-backed open source — my interest is in proving the number, not stating it. Currently looking for full-time SWE / ML-infra / systems roles.
Reflight — Flight recorder for AI agents
Open-source recorder that makes agent failures reproducible: captures
every model and tool call, then replays offline with time, randomness,
and UUIDs pinned. Replays are byte-for-byte identical to the live
run, verified with the network hard-blocked. On a scheduling agent
across 15 gpt-4o-mini runs, a 15-line replay assertion flagged 15/15
violations — an LLM judge caught 1–5 of them at best. reflight promote <run_id> converts any failed run into a CI regression test; replays
finish in ~7 ms at $0.00.
Python · pytest · OpenAI API · LangGraph · OpenTelemetry · Docker
ChainCheck — Real-time hallucination detection · GitHub Action · live
Claim-level detector that flags the exact unsupported sentences instead of scoring the whole response. 94% precision on HaluEval-QA (n=500, gpt-4o-mini judge, balanced 50/50). A 60 ms local NLI pre-filter cuts latency 19× on clear-cut inputs with no precision loss. Ships as a drop-in OpenAI proxy — zero client-code changes.
Python · FastAPI · PyTorch · HuggingFace · OpenAI/Anthropic
40–60% of AWS spend is untagged and invisible to FinOps tools. CostDNA infers per-resource ownership with a GraphSAGE GNN trained on CloudTrail behavior, IAM patterns, and VPC topology, then writes the tags back — pushing existing tools from explaining ~half the bill to ~95%. End-to-end from raw cost report to per-team breakdown in 90 seconds; no data leaves the AWS account.
Caught label leakage in two published cloud datasets (Microsoft Azure 2.6M VMs, Microsoft Philly 117K jobs) by auditing my own results before claiming them — the honest behavioral accuracy is published alongside the inflated first-cut numbers.
Python · PyTorch (GraphSAGE) · Next.js · Terraform · AWS · GPT-4o
Price-time-priority matching engine over a length-prefixed binary TCP
protocol. 0 allocations on the hot path, machine-verified by a
custom global-allocator harness. Cache-padded lock-free MPSC + SPSC
(Acquire/Release memory ordering, // SAFETY: proofs, Miri-validated
in CI). Byte-exact WAL replay + snapshot recovery; group commit
187–245× faster than fsync-per-record at batch=256.
M-series, release, multi-tenant Hub: in-process round-trip ~225 ns · TCP RTT loopback p50 ~78 µs, p99 ~307 µs · matcher walks 1000 levels ~94 µs.
Rust · tokio · Miri · lock-free MPSC/SPSC · CRC32C WAL