/uses
What I build with
The tools, models, and infrastructure I reach for daily. Not aspirational, just what's actually in production.
AI / LLMs
Claude (Opus, Sonnet 4.x)
Primary lane for planning, code review, and agent sessions. Anthropic SDK + prompt caching.
GPT-5 family
Heavy execution and migrations via Codex CLI.
Kimi k3
1M context for whole-repo sweeps, judging other agents' work, and mass fan-out.
Gemini
Vision tasks where Gemini's perception is strongest.
Local models via Ollama
gpt-oss 20B general, qwen3-coder 30B code, qwen3 4B triage, nomic-embed-text embeddings. Runs on a 48GB Mac mini, zero API spend.
Agent Frameworks & Orchestration
Multi-CLI orchestration
Claude Code, Codex CLI, Kimi Code CLI, and Cursor coordinated across the estate, each on the lane it is strongest at.
MCP servers
Tool protocol for code intelligence, shared memory, and semantic RAG across every CLI.
AgentSwarm
Mass parallel fan-out for bulk audits, sweeps, and codegen.
OpenClaw
Multi-agent swarm framework I run my personal automations on.
PairChat
My own local agentic coworker. Delegated grunt work at zero API cost.
GitNexus
Code knowledge graph for blast-radius analysis before any refactor.
Tool calling (native)
JSON-schema tool definitions; preferred over framework abstractions when latency matters.
Frontend
Next.js (App Router)
SSR, server components, server actions. This site, Huadini, blog.
React 18 + TypeScript
Strict mode, no any escape hatches in product code.
React Native (Expo)
Mobile clients for Expired Solutions and StylePilot.
Tailwind CSS + shadcn/ui
Design system as primitives + composition. Avoiding heavy component libraries.
Framer Motion
Used sparingly. Perf cost is real.
Backend & APIs
Python 3.11 + FastAPI
ML inference services, agent runtimes, batch jobs.
Node.js + TypeScript
Edge functions, BFFs, MCP servers.
PostgreSQL + pgvector
Embeddings, RAG indexes, transactional data.
Firebase (Firestore + Auth)
This site's blog, chatbot session store, Huadini auth.
Cosmos DB
Expired Solutions production data store on Azure.
Infrastructure & Cloud
Vercel
This site (Next.js), edge functions, OG image generation, preview deploys.
Azure
Expired Solutions backend (App Service, Cosmos, Blob, ML pipelines).
AWS
StylePilot (Lambda, S3, RDS). Migrated cost down 85% via right-sizing + caching.
Cloudflare
DNS, edge caching, rate limiting.
GitHub Actions
CI/CD, scheduled jobs, deploy gates.
Self-hosted CI on owned hardware
MacBook Air + Mac mini runners for the jobs that should stay on my own machines.
ML / Computer Vision
OpenCV
Pre-/post-processing for produce quality detection.
PyTorch
Custom CV models (Expired Solutions produce classification).
Hugging Face
Pretrained model hub. Embeddings, transcription, segmentation.
CLIP
Embedding-based product/style similarity (StylePilot.ai).
VADER
Fast sentiment analysis when LLM overhead isn't justified.
Azure ML
Training + inference pipelines for production CV.
DevEx & Tooling
Claude Code
Primary IDE-adjacent agent. This site shipped via Claude Code sessions.
Cursor
Inline autocomplete + multi-file edits.
pnpm + Turborepo
Monorepo + fast installs.
Playwright
E2E testing on real product surfaces.
Hardware
Mac mini (M4 Pro, 48GB)
Daily driver. Main dev box, local models via Ollama, always-on agents.
MacBook Air
Runs the self-hosted CI runners.
iPhone 15 Pro
Test device for React Native builds (TestFlight).
Agent skills
Skills Log
The reusable agent skills I write and publish weekly, sanitized and open.
Inspired by the /uses convention.