🤖 Bratan — Self-improving RAG framework driven by an adversarial three-agent closed loop: red-team generates test cases the pipeline fails on, blue-team edits chunking/retrieval/prompts to fix them, judge scores every iteration against a co-evolving test set with Sonnet 4 at temperature 0 (never downgraded). User brings only a corpus + seed cases; blue team owns chunking, hybrid BM25+dense retrieval, reranking, generation, citation verification, and every hyperparameter. Five vector-DB adapters (ChromaDB / Qdrant native-hybrid / Pinecone / Weaviate / pgvector) plus a custom slot. Cost-controlled by a local Qwen pre-judge for inner-loop sweeps + sample-efficient hyperparameter search (Optuna BO, grid, ablation, PSO) — oracle validates every winner. Subprocess-isolated ingest worker, prompt-cache, drift detection, 7 stop reasons. FastAPI + Vite/React setup wizard with on-the-fly seed validation against the corpus, live WebSocket dashboard, GitHub Actions CI. 475+ tests across pytest / vitest / Playwright. Built end-to-end through ~25 sub-agent parallel fan-outs across 6 milestones — the multi-agent build process is itself part of the artifact.
🏎️ f1_RAG — F1 question-answering system with four progressively more sophisticated retrieval architectures (naive → hybrid → +rerank → +router) evaluated side-by-side on a 30-question, 6-bucket frozen eval harness. Hybrid BM25+dense over Qdrant, bge-reranker-v2-m3, Haiku 4.5 routing via tool-use across three retrievers (SQLite stats, Wikipedia narratives, FIA regulation PDFs), Sonnet 4.6 synthesis with structured citations and refusal detection. Self-hosted Langfuse tracing. Built using Claude Subagents in 6 parallel execution waves across 14 milestones — the agentic build process is itself part of the artifact.
🔌 NBA-MCP — Production-style MCP server exposing ~28k NBA player-game rows to Claude Desktop, Cursor, and any MCP-compatible client. Python + FastMCP + asyncpg over Postgres, dual stdio and Streamable-HTTP transports. Demonstrates the "vending a database to AI agents safely" pattern: no execute_sql, every tool a specific parameterized intent.
🔌 n8n — Decision Orchestration Platform A source-agnostic, n8n-orchestrated platform for scalable, secure, resilient, AI-enabled decision automation. This repository is a reference build of an enterprise decisioning platform — the kind of standard an automation/platform team would adopt so that every business function stops reinventing "gather evidence, get an AI opinion, have it checked, apply policy, log it, act" from scratch. The core is n8n as the orchestration engine: every trigger, HTTP call, AI agent invocation, judge call, and database write is a node in an inspectable, versioned n8n workflow graph — not logic buried in application code that only engineers can read. Around that core sit the platform standards enterprises actually need before they'll trust automation with real decisions: API-first and event-driven integration (REST, webhooks, OAuth/JWT), a layered CI/CD gate (build, lint, test-with-coverage, artifact validation, Terraform validation, secret scanning, and an adversarial AI code reviewer), infrastructure-as-code on GCP with a scale-to-zero cost model, observability (structured logging, a Cloud Monitoring dashboard, alert policies), and governance (an auditable decision ledger, least-privilege IAM, policy guardrails that gate every AI-produced recommendation before it can take effect).
🔌 kalshi_soft — Superforecaster Agent for Kalshi "Soft" Markets An autonomous forecasting agent that tests a hypothesis: Opus, equipped with a codified superforecasting methodology (Tetlock / Good Judgment Project), can produce calibrated probability forecasts that beat the market on Kalshi markets driven by human behavior — politics, culture, public statements, behavioral economics/policy — as opposed to stochastic markets (crypto/sports/weather) where research can't out-model the price.
📍 San Francisco Bay Area ✉️ j.allan.wessels@gmail.com