How I Work
I architect systems, define requirements, orchestrate AI-assisted implementation, and validate through testing. This is AI-augmented development — not "letting AI write code," but directing AI with deep architectural understanding to build production systems at speed.
The Orchestration Model
Building the IDP System (44 services, 30 API routers, 56 test files) in under a year as a solo developer required a fundamentally different approach to development. I treat AI as a force multiplier, not a replacement — the architecture, constraints, and quality gates come from me; the implementation velocity comes from AI-assisted coding.
1. Design
Define architecture, data models, API contracts, and constraints. This is the highest-leverage work — AI can't do this well without deep domain understanding.
2. Constrain
Write CLAUDE.md rules that encode architectural decisions as hard constraints. "asyncpg only, no ORM." "Never DROP TABLE if it has data." The AI follows these automatically.
3. Implement
Direct AI-assisted implementation with specific requirements: "Add endpoint X with RBAC check, rate limiting, audit log, tests, and frontend page." Quality checklist is baked into the rules.
4. Validate
Review generated code, run tests, verify in browser. Every PR needs tests. Every endpoint needs a frontend UI. No shortcuts on quality gates.
The Tooling
I use Claude Code with a carefully designed configuration system that makes AI context-aware and consistent across sessions. This is not ad-hoc prompting — it's systematic engineering of AI behavior.
CLAUDE.md — Architectural Constraints as Code
A rules file that every AI session reads. Encodes database conventions, auth patterns, endpoint requirements, and subsystem rules. The AI follows these constraints automatically, producing code that matches the codebase style without manual review of every line.
### Every New Endpoint Must Have
- [ ] RBAC role check
- [ ] Rate limiting
- [ ] Audit log entry
- [ ] Pytest tests (happy path + validation + auth)
- [ ] Frontend UI (page, component, or panel)Custom Agents — Domain-Specific AI Assistants
Specialized agents for different tasks: batch pipeline debugging, migration review, documentation writing. Each agent has focused context and domain knowledge, producing better results than a general-purpose prompt.
Context Budget — Token Efficiency Engineering
A document that maps which files to load for each task type, preventing the AI from wasting context window on irrelevant code. Like a database query optimizer, but for AI context.
Session Handoff — Continuity Across Conversations
A protocol for transferring state between AI sessions. Each session writes a status section that the next session reads — zero re-exploration, instant context loading.
Why This Matters
Backend services in IDP System
Built by a solo developer in under a year. The orchestration approach multiplies output without sacrificing quality.
Test files with comprehensive coverage
AI-assisted doesn't mean untested. The CLAUDE.md rules enforce tests on every endpoint.
AI providers supported
Ollama, OpenAI, Anthropic, Gemini, Groq — with a unified abstraction layer. Architectural thinking enables this flexibility.