Unconstrained conversational LLMs degrade when handling complex multi-file codebases, stateful database queries, and automated refactoring. To eliminate hallucinations and context decay, I architected a Just-in-Time (JIT) modular skill framework for Claude Code and autonomous agents.

Deterministic State Workflows (TDD): Enforced a Red-Green-Refactor state machine that mandates failing unit/integration tests before application logic is generated (-80% regression rate).

In-Database Hybrid RAG (pgvector + BM25): Combined dense embeddings with sparse lexical search via Reciprocal Rank Fusion (RRF), achieving <40ms execution latency on 100k+ records.

AST Tool Governance: Parameter threat scanning, AST regex inspection, and Human-in-the-Loop (HITL) approval gates for sensitive operations.

✔ -65% reduction in manual verification overhead. ✔ -60% token cost optimization via hybrid routing. ✔ 100% deterministic tool calling without schema hallucination.

🔗 Open-Source Repository: https://github.com/yevhens-hue/claude-skills-starter-kit 🌐 Live Portfolio & Simulator: https://shaforostov.pro 📩 Newsletter Deep-Dive: https://yevhenshaforostov.substack.com