Software engineer building backend, distributed and AI systems. MS Software Engineering at San Jose State, graduating December 2026. Looking for full-time software engineering roles starting January 2027.
LinkedIn · Portfolio · Email · acquit on PyPI
NVIDIA (Summer 2026): built an AI-assisted system for Omniverse, NVIDIA's robotics simulation platform, that works out which tests a code change can actually affect, so each change runs only what it needs instead of the full GPU simulation suite. It combines static analysis across Python, C++, containers and generated simulation jobs with deterministic safety checks, so the AI can add evidence but can never skip a test on its own.
Tekion (2022 to 2024): backend engineer on accounts receivable and payments for a dealership platform. Kafka and Elasticsearch reporting, Stripe payment links, a pipeline delivering 4M+ statements a month.
What ties my work together: claims that can be checked. Benchmarks recompute from committed raw data, and when a system misses its own target, the repo says so.
| Project | What it does | Evidence |
|---|---|---|
| acquit | Skips the pytest tests a PR cannot affect, with a re-verified proof for every skip. On PyPI | 198 PRs replayed, 0 unsafe skips |
| benchgrid | Performance test orchestration in Go for scarce, exclusive GPU rigs, with fenced leases | No overlapping runs; the unfenced control duplicates |
| build-sleuth | Triages a failing GitHub Actions run and proposes a patch that only counts if the rebuild passes | 0.915 accuracy vs 0.218 regex baseline |
| tracelab | Tells a real performance regression from a noisy machine | 2,042 real runs |
| Project | What it does | Evidence |
|---|---|---|
| pay-rail | Payment intents with idempotency, an immutable double-entry ledger and a transactional Kafka outbox | 1M fault-injected actions, 0 duplicate ledger effects |
| limit-order-book | Java matching engine: one writer thread, lock-free ring, nothing for the GC to do | 20.0M cmd/sec, 0 GC cycles |
| vertex-kv | Key-value store from scratch in C++17: LSM tree, WAL, Raft with pre-vote | Linearizable reads and writes |
| rail-yard | Job orchestrator in Go with fenced leases and DAG dependencies | 50k decisions replayed, 100% byte match |
| Project | What it does | Evidence |
|---|---|---|
| warpline | Hand-written Triton kernels for LLM inference, profiled with Nsight | 1.52x vs torch.compile |
| agent-serve | One endpoint over three vLLM replicas, pinning each conversation to its cache | 83% prefix-cache hit rate vs 0% |
| pen-stock | LLM gateway in one Go binary: failover, spend caps, OpenAI-compatible | 1.02 ms overhead vs LiteLLM's 12.14 ms |
| medi-agent | Clinical agent platform on FHIR that admits when it has no grounding | Live app |
Open source: 5 merged PRs to RocketRide, an AI pipeline runtime (C++ and Python): runtime delivery fixes, MCP developer tooling, RAG grounding.
Hackathon wins:
- Signet: two sponsor prizes at the DevNetwork API + Cloud + AI Hackathon 2026 (Foxit and Doctavian). Invoices that prove who sent them, checkable with
digandopenssl. - ClaraCare: Deepgram "Voice Operator" prize, DeveloperWeek 2026. A voice companion that calls seniors daily and tracks cognitive change.
- SJ Hopes: 2nd place and Best for Homelessness, SJHacks 2025. Shelter beds on a map for caseworkers.
More builds on Devpost and in projects-hacks.
Certifications: NVIDIA Certified Professional: Agentic AI · NVIDIA Certified Professional: Generative AI LLMs