AI engineer · Los Angeles, open to relocation
I build LLM systems and the evals that show whether they work: a frozen test set, a baseline, the cost and latency of every answer, and the failures written down. There are 37 projects here, each with a live site, and 36 with a write-up.
Most recently I led a natural-language analytics platform on Claude and MCP at Kandu, a stroke-recovery health company. Before that: data scientist at InstaAgent (Y Combinator), and nine years as co-founder and Chief Strategy Officer of an education company that grew enrollment 5.2 times. MSBA at UCLA Anderson, December 2026.
Selected work
01 Full-stack product workshop
SkillCompass
SkillCompass gives two-minute quizzes in more than 40 fields and scores each answer Elo-style against the item's difficulty. You get a percentile against other test-takers, a map of what you know and a study plan. The backend is FastAPI on free tiers, so it costs nothing to keep running.
SiteWrite-upCode02 Agents and MCP road
Marathon Strategy RAG
The same Claude model answers marathon questions three ways: on its own, with retrieval, and with my pace and heat models as tools (also served over MCP). On 42 numeric targets from a question set frozen before any run, the tool coach got 42, retrieval 21 and the bare model 14. A question costs a median 1.7 cents at API prices and takes 8.4 seconds; the write-up shows a full tool-call trace and where each system fails.
SiteWrite-upCode03 Grounded generation language
Podcastify
Podcastify turns a whole book or article into a two-host episode. Each factual line carries the ids of the sentences it rests on and is checked before it is voiced. Across four sources it covered 7.4 of 12 frozen key points against 4.4 for a rule-based script. An independent Opus check found 8 of 172 lines not fully supported, and the repair step did not lower that, which the page says plainly.
SiteWrite-upCode04 Decision tool workshop
Leeway
Leeway plans from the range your car has now, after five years of battery wear, and adjusts it for hills and weather. It keeps a hard reserve at every charger stop and spends a detour budget you set on avoiding unprotected left turns and rail crossings.
SiteWrite-upCode05 Survival analysis track
World-Record Half-Lives
I rebuilt the ratified progression of all 46 Olympic events from World Athletics' records handbook, 1,251 reigns in all, and fitted a survival model with a frailty for each event. Fitted on what was known in 2010, it ranked the records that fell from 2011 to 2025 better than the base rate (Brier 0.235 against 0.329). Duplantis's pole vault leads the LA 2028 watchlist at 48%.
SiteWrite-upCodeHow I test AI systems
Each row comes from the project's own eval files: a test set frozen before the first run, the baseline it has to beat, what an answer costs at API prices and how long it takes. The last column is the failure I would fix next.
| Project | Frozen eval set | Result vs baseline | Cost per task | Latency | Main failure |
|---|---|---|---|---|---|
| Marathon strategy coach (tools, MCP) | 43 questions, 42 numeric targets, frozen by hash before any run | 42 of 42, against 21 for RAG and 14 for the bare model | $0.017 median | 8.4 s median, 14.6 s p95 (n=11) | Asks for a missing input instead of answering; 42 of 42 is close to circular, since the targets come from the models it calls |
| Wine Sommelier RAG | 12 + 30 rubric queries, the 30 committed before the first run | nDCG@10 0.76, against 0.59 for dense retrieval (30 queries) | $0 retrieval; $0.13 generation | Retrieval 196 ms median, 257 ms p95; answer 8.3 s median | Wrong grape or color (20) and an ignored price limit (19) account for half of the 76 missed slots |
| Podcastify | 4 sources, 12 key points each, committed before generation | 7.4 of 12 key points, against 4.4 for rule mode; 8 of 172 lines unsupported | $0.35 per episode (n=1) | 3 min 43 s per episode, half of it audio | The repair step did not lower the unsupported rate, and the two checkers disagree |
| Receipt Auditor (LoRA vs Claude) | 220 items from 158 merchants the model never saw | Fine-tuned 1.5B model 90.5%, against 80.0% for Claude zero-shot | $0 local; Claude $3.03 per 1,000 items | Local 0.19 s median; Claude 3.6 s median, 9.0 s p95 | Hardware and gasoline items, whose training labels themselves split two ways |
| Menu Decoder (LoRA vs Claude) | 303 dishes, held out by dish family | Fine-tuned 69.3% allergen recall, against 79.2% for Claude with the full schema | $0 local; Claude $56 per 1,000 dishes | Local 6.5 s median; Claude 12.3 s median | Rare allergens: sulphites 0 of 11 found. Claude still wins here |
| SkillCompass question bank | 645 shipped questions compared with their drafts | Review pass rewrote 266; longest-option giveaways down from 37% to 23%; key positions rebalanced to 165/164/158/158 | $0 (free tiers) | Not measured | As drafted, 64% of keys were option A. Fixed: options shuffle on every showing and the stored bank is rebalanced |
Cost is the API-equivalent price reported by the Claude CLI; local-model timings are on an Apple M4 laptop. n is given where it is small.
Experience
Jul to Sep 2026
Led the architecture of a natural-language analytics platform on Claude and MCP, from proposal to production: the access model, semantic grounding and promotion gates that let commercial, clinical and operations teams query data themselves. Built the dbt warehouse under it from 13 post-merger systems, under HIPAA and GDPR; identity matching raised record linkage from 75.7% to 97.5%.
Jun to Dec 2026
Designed the Office of Inclusive Excellence's first data pipeline (450+ files, versioned and checksum-verified). Research with Prof. Jane Wu on LA's deep-tech cluster (5,022 firms) and with Prof. Yihao Yuan on how LLM voice agents affect debt recovery across 36,000 calls.
Jun 2025 to Mar 2026
Built the LLM content-generation system for nine small-business clients across X, LinkedIn, Instagram and TikTok, which cut content production cost by more than 80%. Deployed engagement and send-time models that lifted engagement 25% and campaign ROI 18%.
2016 to 2025
Grew from one location to eight with a geo-demographic demand model and A/B pricing tests: enrollment up 5.2 times, revenue 6.7 times, more than 1,000 students. Shipped the company's learning platform end to end, LLM tutor included, and led 28 staff.
Applied AI
Four photo-to-data apps. Each pairs an LLM pipeline with a small Qwen2.5 model I fine-tuned on my laptop (LoRA) for one step of the task, then benchmarks base, fine-tuned and Claude zero-shot on held-out data.
| App | What the small model does | Fine-tuned | Claude zero-shot | Links |
|---|---|---|---|---|
| Receipt Auditor | Spend category for merchants it has never seen | 90.5% | 80.0%a | SiteWrite-up |
| Cellar Scanner | Grape variety from a blind tasting note (1.5B) | 27.7% | 23.0%b | SiteWrite-up |
| Menu Decoder | Cuisine and 14 allergen tags from a bare dish name | 99.3% cuisine, 69.3% allergen recall | 79.2% allergen recallc | SiteWrite-up |
| Race Day Copilot | Km-by-km pacing plan, checked by six rules (3B) | 95.7% valid plans; 0.0% pass all six checks | 90.0% valid; 75.7% pass | SiteWrite-up |
Analysis
Running and wine supply most of the questions.
Wine Score Inflation cellar
Across about 120,000 unique professional reviews (the source file repeats 9,983) nothing scores below 80, and 85% land between 84 and 92. Some critics are simply more generous than others.
Hitting the Wall road
A model trained on 176,059 Boston finishers from 2014 to 2023 reads your splits as you pass each mat and gives a finish range and the chance of hitting the wall. On the unseen 2024 race it missed by 7.9 minutes at halfway, against 18.1 for even pace.
Are Song Lyrics Getting Simpler? language
Fifty years of Billboard hits doubled in length and grew more repetitive, while the vocabulary stayed about where it was.
Index
Each project site has its own look, in one of five families. Pick a family to narrow the list.
About
I spent nine years running a business before I wrote production code. I co-founded Casual Ace Learning Centre in Hong Kong and, as Chief Strategy Officer, grew enrollment 5.2 times to more than 1,000 students across eight locations, and shipped its learning platform, LLM tutor included. That is why my projects report cost and failure modes next to accuracy. I came to AI engineering through an MSBA at UCLA Anderson (GPA 3.92), after graduate analytics coursework at Georgia Tech and a BBA in Accounting and Finance from the University of Hong Kong.
Outside work I have run 14 marathons, with a best of 2:48 in Berlin, and I hold WSET Level 3 in wine. Between them they account for most of the questions on this page.
