Retooling an Engineering Practice Around AI, One Repo at a Time

Wave 1 — build bots for ourselves first (Nov–Dec 2024)

The retooling started with dogfood. In a five-week window the org grew four internal tools: a standup bot, lumo ("a Slack tool that integrated Gmail in Slack" — mine), kangaroo (job scheduling and route optimization for field technicians on the Google Maps API — mine), and frodo, a homegrown GeekBot replacement. None of these were for clients. They were how a services company learns a new stack without billing anyone for the learning curve.

Wave 2 — paid ML work (2025)

Then 2025 fills the org with client-shaped machine learning, and there's a pattern in it: every experiment graduates into a small, boring, deployable API. An ice-cream production forecaster ("XGBoost, LightGBM, Random Forest, DNN, and CNN" with Optuna tuning and walk-forward validation — a colleague's build). A bakery forecasting API I co-authored — FastAPI, engineered lag features, Docker, bearer auth. My Insurate extraction pipeline. Fraud-detection with random forests, an NER model comparison tool, AWS Comprehend research, RAG demos, LLM test harnesses for document generation.

The lesson of wave 2 is in the packaging: the difference between "we played with XGBoost" and "we sell forecasting" is an endpoint with auth and a Dockerfile. Every fad that survived contact with a client got exactly that treatment.

Wave 3 — agents in production (2026)

• February 2026: a spin-off org appears for AI-agent operations, and its repo names are a thesis — agent-memory, agent-guardrails, agent-eval, agent-workflows, agent-gateway. Not prompts: operations.
• June 15, 2026: I shipped Hermes AI Agent as an AWS Marketplace AMI — Packer-built, hardened Ubuntu base, OIDC publishing, and an auto-release workflow that rebuilds when the upstream vendor ships. My favorite commit from that day: "fix: remove em dashes from Marketplace UsageInstructions (rejected by Catalog API)".
• At BackChannel (my other engagement), colleagues run PO/PM agents whose governance README is the sharpest line I've read on agent grounding: "That looked like strategy but was actually past activity dressed up as forward intent." In one knowledge repo, the PM agent has written nearly every entry; humans have one. I run the infrastructure layer underneath.
• And AI got a FinOps discipline: my ThreatMinder work in May 2026 added AI cost-tracking DynamoDB tables, a usage-cost-rates table, and per-product billing tags. Agents that can't be metered can't be sold.

How I keep up: study in public, with receipts

My fork history is a study log with timestamps, and it tracks the work with suspicious precision: llama (April 2024), tesseract (April 2025 — three weeks after starting the Insurate OCR work), crawl4ai and ollama (mid-2025), DocLayout-YOLO (August 2025), langgraph (November 2025), vllm (December 2025), LLMs-from-scratch (March 2026). I fork what I'm about to need, then build something small with it.

• voice-models-benchmark: eight STT/TTS models evaluated on a Jetson Thor — "Results based on real-world tests, not paper benchmarks", down to 97ms first-chunk latency figures.
• claude-telegram-server: Claude Code as a 24/7 Telegram bot, complete with a written postmortem for a 100 MB/sec ArrayBuffer leak that froze the machine in 90 seconds. Side projects deserve postmortems too — that's where the habit gets cheap to practice.
• guide-for-backend-developers: a learning guide I started in 2021, dormant for four and a half years, revived in January 2026 as a "2026 Edition" with an AI Integration section. The guide grew up exactly when the job did.

You don't stay current by reading about the wave. You stay current by shipping something small on every part of it — with auth, a Dockerfile, and ideally a postmortem.

What I would tell anyone doing this

1. Dogfood first. Internal bots are the cheapest tuition a services company can pay for a new stack.
2. Graduate every experiment into a deployable API — endpoint, auth, Docker. That's the line between a demo and a business.
3. When you get to agents, invest in the unglamorous repos: memory, guardrails, evals, cost tracking. The prompt is the smallest part.
4. Keep a public study log. Forks with dates, benchmarks with numbers, side projects with postmortems — it compounds into credibility no certification can match.
5. Meter everything. The first question a client asks about an AI feature is what it costs per call; be the consultancy that already has the table.