State
Typed requests, explicit provider priority, validated output contracts, and review decisions.
Agentic Systems
Agent workflow engineering
Independent projects exploring output validation, failure handling and review policy. The Python reference uses scripted providers; the photography app uses local image analysis and retrieval. Neither calls a live LLM.
Failure handling
Transient retry, permanent fallback, contract rejection, confidence gates, and preserved traces.
Evaluation
Executable hard cases for the primary path, retries, fallback, invalid outputs, review gates, and total failure.
Reliable Agent Workflow reference
A dependency-free Python reference with executable tests for success, failure and human-review paths.
System shape
A small provider protocol keeps model adapters outside the workflow core. The router records every attempt, retries expected transient failures, falls back on permanent or contract failures, validates structured outputs, and marks outputs pending review when confidence is below the request threshold or the provider requests review.
request -> route(primary) -> attempt -> transient failure
-> route(fallback) -> validate contract
-> review gate -> accepted OR pending_review + trace
Executable evidence
The hard-case suite covers primary success, retry recovery, invalid-contract fallback, confidence-based human review, total provider failure, and evaluation reporting.
Photography Mentor with browser-local retrieval
A browser-local multimodal learning application that joins image diagnostics, retrieval, and actionable practice guidance.
System shape
Use local photography knowledge retrieval together with histogram, brightness, color, sharpness, composition, and EXIF analysis. Keep the evidence path visible and translate observations into specific shooting or editing actions.
Engineering signal
Multimodal input handling, local retrieval, deterministic image analysis, constrained coaching output, review state, and privacy-aware browser execution. It is intentionally described as deterministic rather than as an LLM-backed agent.
What remains to be built
The reference validates orchestration policy, not model accuracy. A real provider adapter would need request deadlines, rate-limit backoff, secret management, cost tracking and evaluation on labeled task data. Confidence is supplied by the provider and is not calibrated by this implementation.
Read the review-gate design note or see my production vision experience.