AI application full-stack developer — from model selection and prompting to full-stack delivery, realtime multiplayer, payments/subscriptions, and production operations.
Profile
I build AI that actually runs inside a product: judging, Q&A, recommendations, agent workflows, retrieval-augmented generation, multi-model routing with fallbacks — shipped as production code, not demos.
I own the whole chain: schema design, APIs, admin consoles, auth, payments and subscriptions, deployment and monitoring. The services I run in production are the ones I operate.
I also contribute to public open source: TypeScript / Node / local-first / AI repositories, following the contributor guidelines end to end (issue first, CI green, benchmarks unchanged).
Skills
Languages & frameworks
TypeScript / Node.js / Python
React 19 + Vite
NestJS
PostgreSQL + Drizzle ORM
WebSocket / SSE
Tailwind / shadcn
AI engineering
Structured output with JSON Schema constraints
Function calling and tool orchestration
RAG
Agent workflows
Prompt engineering and eval sets
Multi-provider model routing and degradation
MCP servers
Confidence thresholds with human review fallback
Infrastructure
Cloudflare Workers / D1 / R2 / KV / Pages
Docker / docker-compose
Nginx / OpenResty reverse proxy and TLS
CI/CD (GitHub Actions)
Server operations: backup, rollback, monitoring, alerting
Built from scratch: a 452-puzzle corpus, realtime rooms with state sync, and an LLM judge that answers only 'yes / no / irrelevant / undetermined' with a confidence score.
Judging uses schema-constrained structured output with retry and a logged human-review fallback, so an uncertain verdict is never presented as a conclusion.
Monetisation: full WeChat Pay v3 native flow (signing, callback verification and AES-GCM decryption, query, close, refund) plus a zero-KYC redeem-code path where entitlement is granted in the same transaction as the code redemption.
Runs on a self-managed server (Docker: web / admin / api / jobs / postgres) plus Cloudflare Workers, with monitoring, backup and rollback paths I built.
Turned 'can this be built, how, and for how much' into a public tool: paste a brief, get a feasibility call, architecture and trade-offs, risks, staged milestones with acceptance criteria, and a timeline plus price range.
Lead capture: visitors can hand the brief over with the assessment and contact details, stored in KV and emailed; honeypot plus per-IP rate limiting.
Indexable structure (FAQ, JSON-LD, canonical) so the tool itself works as an acquisition channel.
452 indexable pages exposing only the question text, never the answer; rebuilt daily and pushed to search engines.
Buyer-facing long-form articles (how pricing works, how to pick a vendor, realistic delivery timelines) where every number is filled from live data instead of generic copy.
Multi-source demand radar: two-layer classification (regex plus model) separating 'client looking for a build' from 'peer advertising' and 'full-time roles', with capability-fit scoring, bilingual proposals and a dedicated proposal page per lead.
Market report computed from 440+ live projects; the prices and figures published on my service site come from that dataset.
Six scheduled jobs cover scanning, drafting, the send queue, reply polling, the business report and site rebuilds.
PR to a local-first MCP memory service: separates 'unverified file reference' from 'verified but possibly stale' at render time, with tests and a schema migration.
Followed the contributor path: explained the approach in an issue first, kept CI and benchmarks green locally, then opened the PR.
To be filled in by me (not published)
【待确认】Years of experience and employment history
【待确认】Education and certifications
【待确认】Expected compensation and availability (days per week, remote / on-site / relocation)
【待确认】Contact details (WeChat id, phone)
【待确认】Any prior employer names or attribution requirements