I build AI application/product end to end, solo — from large-scale data pipelines and physics-level simulators to production SaaS with real users, billing, and multilingual RAG.

• DataAssistant AI — AI sales rep for industrial distributors. Live multi-tenant SaaS: Claude + RAG grounded in the customer's own catalog, human-approved replies in 5 languages, Stripe billing. (live product, private code)
• CA-SIM — event-level digital twin of AI data-center power & cooling (grid, load, fault and N-1 events), running entirely in the browser, plus an open event-dictionary spec (EDL). (public demo + open spec)
• TrueProductData — a 935k-product industrial data platform where every spec fact is labelled with where it came from and whether it has been checked, served to AI agents over MCP. (public site, private infra)

davidchen.cch@gmail.com · dataassistant-ai.com · llmdataprotocal.com

Live: app.dataassistant-ai.com · marketing site (EN/ES/FR/TH/zh-TW)

Plugs into a sales team's Gmail or Outlook inbox, scores leads, runs outbound campaigns, and drafts replies grounded in the company's own product / pricing / lead-time data — with a human approval queue in front of every send.

AI engineering (hand-built pipeline, no LangChain):

• Hybrid multilingual retrieval — pgvector semantic search + lexical/SKU matching (pg_bigm), fused with Reciprocal Rank Fusion; Vertex AI multilingual embeddings (chosen because word-boundary-based search fails for Thai/Chinese); locale-aware number normalization so "3,7 kW" survives tokenization.
• A numeric grounding guard — every figure in a generated reply must trace back to a retrieved source row, the customer's own email, or a legitimate derived total (unit price × asked quantity). Anything else is flagged as invented and held for human review instead of sent.
• LLM ops — Claude model tiering by task complexity, per-model rate limiters, Redis response cache, daily cost ceiling with tracked spend, versioned prompt registry with variants.
• Regression-driven RAG — checked-in eval harness (including adversarial/security cases) plus chaos tests: API down, malformed model output, hallucinated figures.

Production engineering:

• Multi-tenant auth (JWT, TOTP 2FA, Google + Microsoft OAuth, team permissions, audit logs), Stripe subscriptions across 5 tiers with quota enforcement at both API and UI layers.
• GCP end to end: 4 Cloud Run services, Cloud Build CI/CD, Secret Manager, Cloud SQL + pgvector, Celery workers with ~12 scheduled jobs, Sentry on both tiers.
• ~110k lines of Python + TypeScript · 300+ API endpoints · 71 DB tables · 30 migrations · 460 commits over 12 months.

Code is private (commercial product) — but the full architecture is documented in dataassistant-ai-architecture, and I'm happy to do a live deep-dive of any subsystem.

Try it: live demo · Open spec: EDL event dictionary (CC-BY 4.0)

For any electrical or cooling event — a grid voltage dip, a source-transfer gap, a feeder fault, a 2 Hz GPU training power cycle, a single device failing (N-1) — CA-SIM answers: how much compute is lost, for how long, and what margin remains. Event-level hybrid engine (dt = 0.5 ms) over SST / BESS / 800 VDC / UPS / CDU / GPU-rack architectures, running entirely in the browser.

• Zero-dependency numerics — ODE integration, Durand–Kerner polynomial root-finding, marching-squares contours, a µs-level protection-window module for fault clearing and selectivity, and OpenUSD export are all hand-written; React is the only runtime dependency.
• EDL (Event Description Language) — an open, versioned dictionary of data-center power/load events with machine-readable waveform profiles and mandatory provenance (grid-code clause, published incident, or public dataset). Certificates cite the exact dictionary version.
• Coverage certificates — parameter sweeps run in a Web Worker into pass / derate / outage heatmaps with zero-margin contours, exported as fingerprinted (sha256) JSON or a print-ready report, plus a diff mode that flags pass→fail flips against a prior certificate. An admissibility rule decides which parameters may be swept on a given diagram, and cells the criterion cannot decide — or that sit on a model-assumption boundary — are labelled as such rather than coloured.
• Grid-code ride-through as data — one registry of 25 public ride-through rules (ERCOT, PJM, CAISO, MISO, NERC, ENTSO-E, EU, UK, AU, OCP, ITIC, SEMI F47, IEC 61000-4-11 …), each with its source clause and its evaluation reading; the simulator reports non-compliance as the architecture's answer instead of tuning the model to pass.
• Calibrated against public data — MIT Supercloud V100 traces and published GenAI H100 power profiles; every model parameter carries a declared calibration grade and every default carries a provenance record. Includes a DC-link control-loop resonance screening module validated against published cases (and the derivation caught a sign error in the source paper).
• Physics under test — 1,000+ tests: a golden-value regression suite treated as a contract (changing a default is changing a golden value), order-invariance and quiescence witnesses, mutation-checked fixes, a statement-exact registry of every "pick the first of several" in the engine, and a validation ladder that computes what may be signed — today nothing is, and the certificate says so. Seven rounds of independent adversarial review are recorded in the repo's design log.

Site: llmdataprotocal.com · 935k products · 7.5M typed spec facts · 220+ manufacturer brands · 160+ categories · served over MCP (private infra)

The pipeline that makes industrial parts data trustworthy enough for AI agents to quote from. Multi-source ingestion (manufacturer sites, sitemaps, APIs, catalogs and PDFs, part-number grammar decoding) → layered normalization → a provenance-first "golden" PostgreSQL database → an MCP server any agent can query.

• Two labels on every fact — where it came from (stated by the manufacturer, parsed from its own description, decoded from the model number, inferred from the product family, or stated by a distributor) and whether it has been checked are stored as independent axes. Over 70% of facts carry a manufacturer URL — product page or document — where a person can open it and see that value; the rest are labelled unverified, not hidden.
• Verification without re-scraping — collectors write one replayable row per page fetched (the URL actually fetched + the raw key/values), so evidence checks run offline against archived corpora: 5.3M facts confirmed against manufacturer pages with zero additional requests. The check is confirm-only by design — string matching can't tell a normalized value from a wrong one, so it never claims to disprove anything.
• Layered, deterministic normalization — the raw layer stores manufacturer wording verbatim and is never overwritten; categories, typed facts and relations are recomputed from it. Claude drafts the mapping dictionaries; the production pass is deterministic and reproducible, so a dictionary fix means recompute, never re-scrape.
• A canonical spec vocabulary as versioned contracts — 254 typed keys following IEC / ETIM / eCl@ss rather than any one vendor: Ie, "operational current / at AC-3 / rated value", and "Rated Operational Current (Ie)" all resolve to one key with unit and qualifiers. Contracts are semver'd and changelogged; a validator diffs every change and requires a version bump when keys are added, removed or re-typed.
• Identity that never drifts — a product is (manufacturer, cleaned part number), enforced by a DB constraint; the part-number normalizer is frozen and corpus-tested, IDs are never reissued or hard-deleted, and merges, splits and brand moves are recorded as aliases. No fuzzy identity matching: the same part number under two brands is common, and is never auto-merged.
• Lifecycle and successors as declared facts — discontinuation evidence is stored per product × region × source, with the manufacturer's own last-order date; "orderable" is only asserted where the manufacturer states it for that part (143k products). 28.7k manufacturer-declared successor and "no replacement" relations; computed cross-brand equivalents are offered with their evidence but never stored as fact.
• Contract-driven MCP server — category-agnostic tools (search_products with parametric spec constraints, find_equivalents for cross-brand substitutes, get_datasheet, …) driven purely by the data contracts, returning each value with its provenance label and, where verified, its evidence URL: a new product category becomes agent-queryable with zero new code.

AI systems: Claude API (structured outputs, tool use, model tiering, cost governance) · RAG (pgvector, hybrid retrieval, RRF, eval harnesses) · MCP servers · prompt registries Backend: Python · FastAPI · SQLAlchemy 2.0 async · PostgreSQL (pgvector, pg_bigm) · Celery + Redis · Alembic Frontend: React 18 · TypeScript · Vite · Tailwind · Redux Toolkit / react-query · i18next (5 locales) Infra: GCP (Cloud Run, Cloud SQL, Cloud Build, Secret Manager, Vertex AI) · Docker · Stripe · OAuth (Google & Microsoft) · Sentry · GitHub Actions Simulation: hand-rolled numerical methods · DSL design · OpenUSD · Web Workers

Building AI products and want to compare notes — or hiring for a team that ships? Email me: davidchen.cch@gmail.com