I take ambiguous problems from data and architecture through deployment, testing,
and failure analysis, from parcel-level geospatial products to low-latency C++ infrastructure.
Parcel-level site-feasibility for Woodburn, Oregon, on real Marion County data.
Owned end to end: data, architecture, client, API, deployment, and tests.
ProblemParcel feasibility means combining fragmented zoning, hazard, utility, and geometry data into one buildable-area answer.
What shippedThe React/MapLibre client, the PostGIS feasibility API, a real-data ingestion pipeline, the deployment stack (Vercel plus an Oracle Cloud VM behind a Cloudflare tunnel), and validation tests.
Engineering judgmentI union overlapping constraints before subtracting so nothing double-counts, and I show the browser's planar answer (51%) against the API's geodesic one (50.66%) instead of hiding the gap. An evaluation harness caught an unreliable LLM extraction approach before production: the first real Cerebras run scored 1 of 6 against a hand-verified gold set, which I treat as useful failure detection, not an AI accuracy metric.
A career workspace that runs entirely on your own machine and only writes résumé claims it
can trace back to verified evidence — built and governed with an AI-assisted engineering
workflow of its own.
ProblemJob search tools make it easy to generate résumé claims that sound right and don't hold up under a technical interview. I wanted a system that would refuse to do that by construction, not by good intentions.
What shippedA Next.js and FastAPI application that runs entirely on your own machine, pulling public job postings from Greenhouse and Lever, holding a versioned record of verified career evidence, and only generating résumé material that traces back to that evidence, with human review and DOCX export as the last, explicit step. Built with Claude Code inside a governed workflow — bounded work packages, isolated worktrees, independent review — rather than as an unsupervised generator.
Engineering judgmentRetrieved job postings and model output are treated as untrusted data, not instructions: nothing an LLM produces can grant itself tool, network, credential, or send authority, and no generated claim reaches an export without human review. The governed workflow caught real defects, not hypothetical ones. An independent fresh-context review found that the tests meant to prove no unintended content reached an exported document were asserting over compressed bytes, so they would have passed either way, and it proved that by breaking the check and watching them stay green. Separately, a test I had written off as a flake turned out to be a real Windows race in the work-package lease tool: a killed process keeps its byte-range lock for another 1 to 3 ms of kernel rundown, so the lease refused a holder that was already dead. Both fixes are pinned by mutation tests rather than a weakened assertion.
Private repository — this covers the architecture and governance approach directly; happy to walk through the code live.
Roguemouse
applied AI · governance
A multi-agent AI operations and governance system built for Milan AI Week 2026.
Public and MIT-licensed.
Three reasoning roles, a Risk Officer, an Ops Engineer, and a Synthesizer, that produce a structured proposal or an explicit refusal, never free-form output taken on faith.
Every decision writes a SHA-256 hash-chained audit record to content-addressed, S3-compatible object storage; the chain is verifiable client-side.
Explicit degraded-input handling and a 327-test baseline in CI.
Built for Milan AI Week 2026; the hosted deployment was retired after judging.
Low-Latency Order Execution Engine
systems depth · c++
A lock-free matching engine, designed and benchmarked for bursty order flow,
with every memory-ordering choice justified in the code and every number reproducible.
Lock-free SPSC queue with alignas(64) cache-line padding to kill false sharing, benchmarked at a median ~103M ops/sec. The MPMC variant uses per-slot sequence numbers with CAS retry for ABA safety.
Price-time-priority matching with partial fills and indexed price-level lookup for cancellation over FIFO order storage per level, at ~0.9 µs p50 / 2.1–2.3 µs p99 order-to-fill across 100K orders.
A Clang ThreadSanitizer run exposed a lost-wakeup deadlock in shutdown (the stop flag was set outside the condition variable's mutex); I fixed the mutex discipline and added a regression test. 11 of 11 tests pass, incl. a 50-thread deadlock torture and a 1,000,000-item MPMC stress with zero lost.
order-to-fill latencymodeled
warming up…
Deterministic synthetic distribution calibrated to the engine's measured p50 ≈ 0.9 µs and p99 ≈ 2.2 µs (i5-12600K, MSVC Release). Not raw benchmark samples or live telemetry.
Real-time simulator of why bid-ask spreads exist: an adaptive market maker that
detects toxic flow and widens spreads, visualizing the Glosten-Milgrom model.
Diagnosed TCP/IP, DNS, and VLAN faults across 15–20 production client environments; resolved 80–90% of incidents remotely through protocol-layer isolation.
Wrote stateful firewall rules and network-access-control policies on production Linux.
Delivered root-cause analysis to non-technical clients across 15+ accounts, translating VLAN and DNS failures into actionable remediation steps.
Software Engineering InternT. Mann FinancialSummer 2023
Built Python data pipelines that segmented platform usage by cohort; the resulting drop-off analysis informed subsequent product changes.
Implemented JavaScript event tracking that enabled campaign-level engagement analysis and data-driven targeting.
03
How I work
Own the whole problem
I move from unclear requirements through implementation, deployment, instrumentation, and documentation.
Make tradeoffs visible
I publish failure cases, limitations, benchmark environments, and the evidence behind technical claims.
Use AI with verification
I build evaluation harnesses, structured outputs, refusal paths, and deterministic checks instead of treating model output as truth.
