Weekly entries from a six-month AI product development program.
Built a multi-agent adversarial substantiation workflow that pressure-tests marketing claims through parallel evidence gathering and an attack-rebuttal loop, then measured the cost delta between plan-in-context and plan-in-script orchestration.
dynamic-workflows, multi-agent, claims-governance, cost-analysis, mcp
Built the capability surface for Claims Desk — a 4-tool MCP server, Supabase-backed registry seeded with 12 test claims, and two Agent Skills, verified end-to-end across Claude Code and claude.ai.
mcp, agent-skills, supabase, railway, python, claim-governance, llm
A retrospective on building richardthomchick.com from a static MVP through a four-phase evolution spanning unified search, AI-powered content interactions, visual animations, and evaluation infrastructure.
astro, tailwind-css, pinecone, openai, anthropic, rag, semantic-search, pagefind, vercel, ai-evaluation, text-to-speech, server-sent-events, canvas-animation
Built four distinct product modules on a solid data foundation, then discovered the gap between architecturally correct and user-ready through a design critique.
advisor, rag, retrieval, llm, product-design
Built the information architecture for a B2B buying group personalization advisor: a canonical data model and a nine-document corpus designed for both human practitioners and AI retrieval.
information-architecture, rag, data-modeling, knowledge-base-design, retrieval
Built Dino, a consumer-facing AI dining concierge for Las Vegas with Rat Pack personality, real Google Maps restaurant data, mock reservation booking, and Google Calendar deep links — deployed on Railway via a FastAPI agentic architecture.
agentic-architecture, fastapi, claude-code, conversational-ai, google-maps-api, railway, consumer-product, personality-engineering, tool-use
Built a conversational AI intake copilot that transforms messy stakeholder feature requests into structured specs using a two-gate review architecture.
conversational-ai, claude, streamlit, supabase, safe, product-management, advisor-tool, multi-call-llm
Took the SAFe Feature Spec System from prototype to production by wiring in governance, migrating from SQLite to PostgreSQL, building a ConnectorInterface abstraction, and deploying the full v3 system to Streamlit Cloud.
advisor-tool, postgresql, streamlit, python, safe, evaluation, production-deployment, connector-pattern, governance, anthropic-api, supabase
Built nine responsible AI modules (cost guardrails, grounding checks, content safety, bias detection, audit trails, and prompt governance), transforming the SAFe Feature Spec System into a pipeline that can be trusted in production.
responsible-ai, guardrails, grounding, content-safety, bias-detection, audit-trail, cost-governance, prompt-governance, streamlit, sqlite, python, claude-code
Built an evaluation pipeline for the SAFe Feature Spec System: SQLite persistence, prompt versioning, a golden test set, a Streamlit dashboard, and an AI improvement suggester, turning prompt engineering from guesswork into measurement.
evaluation, llm-as-judge, sqlite, streamlit, prompt-engineering, a-b-testing, python, claude-code, safe, prompt-versioning
Built a six-agent Streamlit system that automates the full SAFe feature spec workflow — from classification to scoring to polish — replacing a multi-session manual process with a consistent, 10-15 minute pipeline.
multi-agent, streamlit, safe, feature-spec, session-state, reflexive-architecture, llm-as-judge, workflow-automation
Upgraded the Knowledge Assistant from a local ChromaDB prototype to a production-grade Pinecone-backed system, added agentic retrieval, and refactored tools into an MCP-style composable layer.
rag, pinecone, agentic-rag, mcp, vector-database, tool-use, streamlit, multi-agent, embeddings
Built a full RAG pipeline from embeddings to deployed Knowledge Assistant, explored hybrid search and re-ranking, and established a baseline evaluation framework for the Feature Spec Generator.
rag, embeddings, chromadb, vector-store, hybrid-search, re-ranking, metadata-filtering, streamlit, evaluation, llm-as-judge, openai
Converted the Feature Spec Generator from a CLI tool into a Streamlit web app, deployed it to Streamlit Cloud, and shipped it to real users with Teams webhook notifications.
streamlit, deployment, webhooks, microsoft-teams, session-state, feature-spec-generator, production
A grueling but rewarding week building production-grade multi-agent systems — from a hierarchical Feature Spec Generator to an ROI Analyzer hardened with resilience patterns, cost optimization, and smart model routing.
multi-agent, tool-use, function-calling, resilience, cost-optimization, prompt-engineering, asyncio, model-routing
Built a progression from calculator tool to autonomous research agent, and learned that resilience code, context management, and bounded workflows aren't optional — they're load-bearing.
tool-use, agents, python, context-window, prompt-engineering, error-handling
Hands-on experiments with tokens, context windows, streaming, and prompting techniques reveal that AI product management requires an entirely new economic mental model.
tokens, context-window, prompt-engineering, streaming, python, rate-limits, product-management
A non-engineer sets up a Python dev environment, generates an API key, and makes a first Claude API call — and discovers that vibe-coding has limits.
python, anthropic-sdk, dev-environment, cli, api