I run a small fleet of AI agents in public, and publish what they cost.
Product manager building, measuring and governing agentic workflows. This page is the register: what is planned, what is running, what it costs per verified outcome, and what I decided not to automate.
NOW · B10 Evaluation Harness. The Foundation · week 2 of 54
PERMISSION · no rung reached yet. Nothing I have built acts without a human. Next: suggests, gated on B11.
The run ledger starts writing here when Surface Zero goes live in B11. Until then these are empty on purpose.
What these agents are allowed to do
fig. 1 · consequence ladderEvery rung is a permission a third party can revoke, and each one is gated on a block landing first. A dashed rung has not been reached — until it is, the work behind it is described as implemented, never as operated.
Built and published
full register →B8 · 2 live artifacts
Smart Reply Copilot: Multi-Step Prompt Chaining
Built an AI-powered customer support assistant that routes tickets through a 3-step classification chain: category detection → tone calibration → draft response generation. Each step includes confidence scoring for transparency and human-in-the-loop review.
B7 · 3 live artifacts
AI ROI Calculator: Translating Technical Metrics into Business Value
Built a financial modeling tool that bridges the gap between technical execution and executive decision-making. Translates AI system costs into CFO-ready 3-year projections using real production telemetry from Block 6. Demonstrates the critical PM skill of business value communication—turning 'it works technically' into '$1.17M annual value' with cited, defensible assumptions.
Everything else on the shelf
7 blocksFoundations: API & Prompt Engineering
Mastered the fundamentals of LLM interaction, moving from 'Chat' to 'API'. Learned to control output structures via System Prompts.
Help Desk Copilot: RAG-based Q&A System
Built a conversational AI assistant that answers questions about support ticket data using complete RAG implementation (Retrieval-Augmented Generation). Goes beyond keyword search to synthesize answers, identify patterns, and provide insights across the entire corpus—demonstrating the same architecture powering ChatGPT and modern AI assistants.
Full-Stack Prompt Engineering & Governance
Evolved from basic scripts to a production-grade Web Application. Built a 'Prompt Playground' that operationalizes prompt templates and enforces ethical guardrails via UI logic.
Fine-Tuning & Local Inference
Pivoted to 'Small Language Models'. Fine-tuned Llama-3 to perform niche tasks cheaply, proving that bigger isn't always better.
AI Product Owner: Agentic Backlog Generation
A CLI agent that ingests raw support tickets from a CSV and writes developer-ready user stories, using a ReAct loop with a reflexion pass for quality control and a Zod schema to force INVEST-compliant output. Adds a dry-run cost estimate and RICE scoring, then wraps the CLI in an Express API.
Voice of Customer: Unsupervised Insight Engine
Built an AI-powered analytics platform that transforms unstructured support tickets into strategic insights using unsupervised learning. Automatically clusters thousands of tickets, generates executive summaries, and surfaces hidden patterns—eliminating hundreds of hours of manual categorization while maintaining strict data governance and cost controls.
Market Intelligence Agent: Hybrid RAG with Autonomous Routing
Built an autonomous AI agent that solves the 'stale data' problem by intelligently routing queries between internal knowledge (vector store) and live web search (Tavily API). Demonstrates advanced RAG architecture with self-routing capabilities—the agent autonomously decides which data source to query based on the question's intent, eliminating training data cutoff limitations.