Software engineer at UF Health. I write the pieces most teams import — training loops, retrieval, agents, inference serving, consensus, gateways — from scratch, test them in CI, and measure them on real hardware. Ten open-source systems below. Every number comes with the command that produced it.
Each repository is small enough to read in an afternoon and complete enough to run: a test suite, CI, a benchmark table with the command that produced it, and a roadmap that says what is still missing. All measured on an 8 GB laptop with no GPU.
A GPT trained end to end on a laptop CPU: a byte-level BPE tokenizer written from the merge rule, explicit causal attention with a tested fused path, and KV-cache generation that is proven identical to recomputation.
Raft from the paper's Figure 2 and a linearizable key-value store with ReadIndex reads and exactly-once client sessions, tested by an adversarial network simulator under the race detector.
An OpenAI-compatible gateway in front of OpenAI and Anthropic: virtual keys, per-key rate limits, retries with jitter, provider fallback, circuit breakers, a deterministic-request cache, streaming translation and cost accounting.
Hybrid retrieval-augmented generation in one SQLite file: a from-scratch BM25, a deterministic hashing embedder, reciprocal-rank fusion, reranking, cited answers, a CLI, an API, and an eval loop that judges every change against a measured noise floor.
The core of an agent framework in 1,400 lines: typed tools with generated JSON schemas, a budgeted loop, parallel tool calls, validated structured output, JSONL traces and a Claude provider tested against a mocked SDK.
An eval harness for LLM apps: declarative tasks, nine metrics plus LLM-as-judge, a response cache, bootstrap confidence intervals and a paired comparison that finds changes overlapping intervals would hide.
An inference server with dynamic batching, bounded queues that answer 503 with Retry-After, graceful drain, readiness probes, Prometheus metrics and a Docker image, with a load test that measures the gain.
A PyTorch-shaped tensor library with reverse-mode autodiff over broadcasting NumPy ops, Conv2d, LayerNorm, Adam and schedulers. Every gradient of all 30 ops is checked against finite differences.
The index under every vector database, from the paper, in dependency-free Go: layered graph, beam search, heuristic neighbour selection, tombstone deletes, CRC-checked persistence and one allocation per query.
A sandboxed Model Context Protocol server: read-only SQLite through a token-level SQL gate, an allow-listed filesystem with realpath containment, and allow-listed HTTP fetch, over stdio and Streamable HTTP, with zero native dependencies.
From an IRB-governed research platform to a zero-to-one video engine to enterprise ad tech: scoping ambiguous problems with stakeholders, owning architecture through rollout, and handing off systems that keep running.
The habits behind the repositories above. They are the same habits I bring to a team.
Every benchmark names the machine, the command and the noise floor. If a change does not move a number, it does not ship. If it moves a number the wrong way, that is reported too.
Tests and demos run without API keys or network, against deterministic fakes, so CI is honest, fast and reproducible by anyone who clones the repo.
Raft from Figure 2. HNSW from Malkov and Yashunin. BPE from the merge rule. Frameworks are easier to use well once you have built what they hide.
A platform for 100+ researchers, a serverless engine for 10,000 accounts, an app on the App Store: discovery to rollout to documentation to handoff.
A voice-agent platform for regulated industries where agent actions are verifiable by design: a compliance finite-state machine constrains behaviour in real time, and a cryptographic audit ledger makes every action tamper-evident.
An AI news summariser that turns jargon into plain language. Cross-platform MVP shipped in 48 hours; production RAG backend on pgvector and Prisma, served on an AWS Bedrock and SageMaker pipeline.
A menu-bar app for per-app volume, multi-device output, audio routing and a 10-band EQ. A free, open-source alternative to SoundSource.