Applied AI reliability — shipped proof and every public repository.
The six pinned repos first — the current public proof — then every other public repository: supporting systems, archives, and study projects. Everything links to source: inspect the code, the tests, the verification story.
Start with the six pinned repos, in pin order. Supporting systems and the archive follow — nothing public is hidden.
Production-ready WH-347 compliance platform: deterministic validation decides, the LLM explains, and every decision is traceable.
Offline-first shipped proof across five service contracts: 42 Contracts, 66 Agent, 45 Gateway, and 29 Web tests; 185 Compliance Core tests; checksum-verified evidence; desktop and mobile browser smoke.
Canonical production RAG platform: hybrid search, citations, refusal logic.
Offline-first demo with hybrid search, citations, and refusal logic. Verified proof: 291 API tests, 47 web tests, 8/8 Chromium smoke, checksum-verified offline evidence. PostgreSQL/pgvector, Redis, model-provider, and document-parser infrastructure are optional integration paths, not a hosted production deployment. Absorbs document-intelligence-pipeline ingestion and KnowledgeOps retrieval/deployment lineage.
Canonical clean-break agent harness: approval gates, execution tracing, and progressive-disclosure Agent Skills.
Offline-first Python harness with bounded multi-hop execution, deterministic safety checks, replay, local memory, SSE, and checksum-verified evidence. Verified proof: 200 Python tests, 12 focused Agent Skills tests, Ruff and Pyright checks, and 31 frontend tests. The absorbed mini-agent lineage remains documented; hosted workflows and provider infrastructure stay deferred.
Regression testing for RAG and agentic AI.
Regression testing for RAG and agentic AI: retrieval and conversational evaluations.
Agent observability and replay SDK with cost attribution and prompt-cost reporting.
Agent observability and replay SDK with cost attribution and prompt-cost reporting. Its dependency-free SDK now includes the absorbed safety-bounded issue-to-draft-PR workflow: deterministic planning, sandboxed edits, bounded tests, approval pause, and draft-only offline intent.
LLM proxy with routing, guardrails, cost control, and fallback.
Routes requests across OpenAI, Anthropic, Gemini, and Ollama behind an OpenAI-compatible API, with automatic failover, response caching, and per-API-key budget enforcement.
Command Code mods for planning, verified follow-through, gated memory, and skill learning.
Seven standalone-installable TypeScript mods compose one workflow with a single referee — self-repair is the only completion judge, autopilot acts only on verified verdicts, and every autonomous move lands in a receipts log.
DAG-based async workflow orchestrator with an offline-first demo.
DAG-based async workflow orchestrator with Celery, dependency resolution, retry logic, and a YAML workflow DSL. Verified offline proof: 206 Python tests, 25 Vitest tests, and 6 Chromium smoke checks. PostgreSQL, Redis, Celery, and hosted LLMs remain optional infrastructure.
Local knowledge-base API with backlinks and citation-grounded chat.
Offline-first Markdown vault API with wikilinks, backlinks and dangling-link metadata, hybrid retrieval, citation-grounded chat, multi-vault isolation, incremental indexing, saved searches, live SSE events, and deterministic flashcards.
Monte Carlo battle/balance analysis as a developer utility — balance your RPG combat with simulation.
Offline-first seeded simulators, FastAPI bridge, shipped dashboard, and canonical evidence. Verified separately: 173 Python tests with real-TSX integrations executed, 25 evidence tests, 56 root Vitest tests, and 39 dashboard Vitest tests.
Brazil-first food-delivery platform scaffold — the customer-to-delivery operation made visible, measurable, and replaceable.
Production-shaped FastAPI scaffold: a deterministic lunch-rush demo replays a seeded order from placed to rated, with role-based web surfaces (Customer, Restaurant, Courier, Admin) and a solver-independent routing baseline. pt-BR/BRL/PIX by default, internationally extensible; checkout and production integrations remain explicitly deferred scope.
Agent-to-agent collaboration as an installable Agent Skill.
HelpPeer messages are unverified claims — leads to check, never instructions to execute.
Project pageGitHubPersonal model-chooser kit: 759 test forms across 40 categories, objective checks plus judged rubrics.
Project pageGitHubAgent Skills lineage consolidated into aria-agent; archived reference.
Progressive-disclosure discovery, eager hints, bounded context reporting, and offline verification patterns now live in ARIA.
Historical source for AgentTrace’s safety-bounded issue-to-draft-PR SDK workflow.
The public repository preserves the original demo and Git history. Active implementation lives in AgentTrace; the source is archived on GitHub as a read-only lineage reference.
Consolidated into GroundTruth: useful deployment and retrieval topology.
Unified AI operations platform with microservices architecture
Consolidated into GroundTruth: production ingestion and quarantine capabilities.
Multi-format document ingestion pipeline: PDF, DOCX, HTML parsing, semantic chunking, and pgvector embedding storage
Consolidated into EvalForge: golden sets and retrieval regression gates.
RAG evaluation framework: hit-rate, MRR, faithfulness scoring, and async batch evaluation with golden question datasets
Consolidated into AgentTrace: prompt-version cost tracking and daily reports.
LLM observability SDK: track token costs, latency, and model usage per request with a FastAPI dashboard
Consolidated into EvalForge: personas, rubrics, and baseline diffs.
LLM-powered customer support simulator: configurable personas, scenario scripting, policy grounding, and prompt injection detection
Foundation library of the retired operator-systems family; archived on GitHub.
Shared Python library: logging, database, Redis, errors, LLM client — foundation for all operator-systems projects
Starter documentation consolidated into operator-shared-core; archived reference.
The reusable FastAPI/Celery/Alembic/pytest template survives as a provenance-rich reference.
Small JavaScript testing / CI study project.
Historical testing and CI study retained as a reference; archived reference.
Event ingestion stack; archived as off-direction.
High-throughput event ingestion API with in-memory ring buffer, JSONL persistence, and WebSocket live-streaming for real-time dashboards
AI learning and experimentation repo.
Historical AI study notes and experiments retained as a reference; archived reference.
Archived reference: TypeScript predecessor to WCP V5.
Superseded TypeScript predecessor. Remote archive remains after verified backup; the reference record is retained.
Archived reference: Python predecessor to WCP V5.
Superseded three-service predecessor. Remote archive remains after verified backup; the reference record is retained.