Data Engineering · AI Evaluation
I build data pipelines that show which step lost or rejected data, and evaluation gates that check AI answers against their evidence. When evidence is missing, the answer is withheld.
Airflow · PostgreSQL/TimescaleDB · pgvector · Individual research
Loads public NASA IMS bearing-vibration data into TimescaleDB only after QA, idempotently and with lineage; failing data is quarantined. A maintenance report is issued only if citation hashes, entailment, and equipment scope all verify; otherwise ABSTAIN.
Boundary: not a production system. Anomaly-detector performance is INSUFFICIENT_EVIDENCE, so no detection claim. Public CI skips database tests; the full PostgreSQL run is a manual release gate.
FastAPI · PostgreSQL/pgvector · React · Individual project
Shows Samsung Electronics and SK hynix OpenDART figures with filing evidence; an answer is confirmed only if company, period, report, and consolidated/separate basis match the query and the latest valid filing is used.
Boundary: the public demo has ended; PostgreSQL integration tests were not run in public release QA. No operations, traffic, or performance claim.
FastAPI · PostgreSQL/pgvector · SSE · Individual project
A document Q&A API that searches only uploaded files and abstains without a relevant passage. Each upload is one PostgreSQL transaction, same-filename uploads are serialized, answers are not cached if any document changed during generation, and streaming never retries once answer text has been sent.
Boundary: not a production deployment. Known limits, including a possible deadlock, are listed in How it works.
Product Engineer Intern, Factagora · Sep–Nov 2026. Evaluation tooling for an LLM evidence-classification API. My contributions: a counter for responses filled with defaults instead of analysis, and a replay that regenerates test results to compare with the saved ones.
DB and HR Manager, proptech startup · 2023–2025. Automated one cross-team handoff with Slack, Google Calendar, and Google Sheets and timed it. The before/after timing is self-recorded, not an independent benchmark.
Each README states its own limits.
- Lease renewal baseline · deployed service comparing a jeonse (deposit lease) renewal offer with actual renewals in the same complex and floor-area band.
- Quant Research Observatory · static site for an ETF momentum backtest and its verification evidence; the research code is private.
- PUBG Telemetry Watch · pre-registered review of rule-based anomaly signals in match telemetry, reporting a null result.
- Agricultural weather pipeline · five-person capstone; I owned collection from three public APIs and Bronze-to-Silver cleaning.
- Resumable collector · individual project: file-based resume state and idempotent UPSERT.
- Three-tier observability · five-person course assignment; I rebuilt my CloudWatch and k6 part in 2026.
Responsible AI · Knowledge graph · Individual entry
Turns 21,386 public UK Modern Slavery Registry records into a citation-backed human review queue. Team registration was required, so I entered as a one-person team. Flagging calls no LLM: records are linked across years only by an exact, nonblank company number; name-only or conflicting identities stay on HOLD. Live review workbench
Recognition: Top Ten Honors; later invited to the founding cohort of Austin AI Hub's six-month AI Hub Incubator.
Boundary: review questions, never accusations; not a production system. The invitation implies no program completion or pilot deployment, and "individual entry" means no human teammates, not AI-free coding.
TypeScript · Google ADK / Gemini · Solana Devnet · Prototype
Gemini recommends products but holds no payment authority; three buyers' payments run as one atomic Solana transaction only after human confirmation and deterministic budget, product, expiry, and cost-share checks. Demo and evidence
Boundary: atomicity does not cover Firestore and Solana as one distributed transaction. One operator confirms all buyer roles on Devnet test tokens, so independent wallet approvals and payment safety are unproven.
Codex plugin · Typed invariants · SHA-256 archive
A reversible context-control plane for long Codex tasks: typed invariants protect exact constraints, sanitized evidence is keyed by SHA-256 hash, and transitions stop when a checkpoint fails integrity checks. Evidence explorer
Boundary: SHA-256 gives content addressing and integrity checks, not encryption or authenticity. "Reversible" means restoring a verified checkpoint and its retained evidence, not Git state, files, commands, or side effects. A synthetic savings benchmark failed, so no cost-saving claim.
Node.js · GitHub Copilot SDK · Azure App Service · Individual entry
A productivity app where a GitHub Copilot SDK agent orders the day's tasks and drafts the top-priority one. Recorded 6th of 39 entries on the LipCoding 2026 leaderboard; not a formal award.
Boundary: Azure resources were deleted after the event; if Copilot sign-in (token or CLI login) or the CLI runtime is unavailable, the app uses a deterministic fallback.
I build with Claude Code, Codex, and GitHub Copilot, using them to compare implementations and generate candidate tests. I accept a change only after an invariant check, a diff review, and negative-path tests, and narrow the public claim when evidence falls short.