The useful question is not whether a demo looks intelligent. It is what the system can do, what it is allowed to change and how failure is handled.
Public personal engineering
Agent evaluation and orchestration
A Python evaluation runtime for Agent Skills. A dependency graph schedules evaluators, joins their evidence and gates whether to accept, revise or escalate.
- Python
- DAG scheduling
- Evaluation
- Staged repair
Read the engineering notes for SEG ↗ Working prototype
Controls around model output
A Python CLI for policy retrieval, deterministic checks and optional model analysis. Unsupported citations fail closed; approval stays with a human.
- Python
- OpenAI Responses API
- Citation validation
- Human approval
Read the engineering notes for EvidenceGate ↗ Internal product
Product knowledge with a deliberate access boundary
Python retrieval from a real Seller account and a separate Semantic Kernel RAG application, using vector embeddings in Cosmos DB. Material changes stayed manual.
- Python
- C#/.NET
- Semantic Kernel
- Cosmos DB
Read the engineering notes for Seller RAG ↗