Justin Smith

Product leader who ships AI systems

I have worked on AI products for close to a decade, and since 2025 I build them as well as direct them. An AI bidding engine at Acquisio in 2016. Twelve AI startups at Creative Destruction Lab alongside Mila in 2017. Two ML models shipped to production at Officevibe in 2019. The MLOps process behind a second production team at Workhuman in 2021. Then a year building it all myself: 73 unattended jobs, roughly 90 agents, and a 48-agent product organization.

~10 yrs shipping AI products
73 scheduled jobs running unattended
11% → 31% activation in 4 months at Workhuman

The unusual part

I ran product at a talent-assessment company, so when I received my own psychometric profile I treated it as an engineering spec. I score 95 on Accountability and 13 on Establishing Order: I feel responsible for everything and naturally sequence nothing. So every agent in my system forces external capture on loose talk: owner, deadline, next action. Each score tension became a system requirement. The assessment is not a report I read once. It is the operating system.

Start here

The Machine

The full essay. One operator, seventy-three unattended jobs, roughly ninety agents: what I built, what it taught me, and what broke along the way. Seven chapters, about fifteen minutes.

Case studies

The Assessment That Became an Operating System

How psychometric score tensions became the design constraints for an agent fleet, and why the pure-LLM first version had to die.

Every Agent Wakes Up Knowing What the Others Did

Cross-agent persistent memory: the architecture, and the five failures that taught me memory is a distributed-systems problem in a productivity costume.

One Sentence In, One Company Out

One logged sentence to a shipping MVP: overnight research, then a 29-issue product program run by agents through two phase gates, then React and Vite code deployed on Vercel.

Cheap Models, Expensive Mistakes

A 20-trial harness measuring where local models can replace frontier ones. The local agent did not hallucinate, it silently did nothing 20% of the time, and that shape of failure decided the routing policy.

I Built the Product Org I Used to Hire

A 48-agent product organization with a Chief Product Officer agent. It ran a 29-issue gated product program for PsychBill end to end, and the findings became committed code the next day. What it revealed: craft specifies, judgment does not.

Demos

This renders from a script rather than a screen recording, so it regenerates against the current state of the machine instead of aging into a claim I can no longer reproduce.

Terminal: job counts, then a health audit reporting 11 findings and 28 passing checks.

Platform health: nine check modules over the whole machine in about six seconds. Uncurated, including two bugs the capture exposed in the reporter itself.

Systems running right now

The full architecture map lives at github.com/JustinTSmith/systems. The short version:

Memory spine

Every AI session starts knowing what previous sessions learned. An overnight job briefs each agent on what the others did.

Product Compass

A 48-agent product organization: a CPO agent routing to eight VPs and Directors, with roughly forty specialists beneath them.

Agent fleet

One Telegram thread into a three-agent fleet, plus 13 agent companies on Paperclip running roughly ninety agents.

Voice pipeline

One press of the iPhone Action Button: transcribed, classified 8 ways, filed into the vault in seconds.

Research to audio

Four parallel weekly research scans become a written briefing and a podcast-style audio overview I listen to instead of read.

Self-healing layer

Watchdogs restart dead services, nightly security review, health checks that commit their own reports.

Agent companies

13 companies and roughly 90 agents on Paperclip: PsychBill, a medical council, an equity research desk, staffed professional teams.

Local and cloud routing

Ollama and MLX serve local models for mechanical work; judgment escalates to frontier models. One pipeline analyzes on a local 35B and writes on Opus.

Shipped and open

AI and ML in production, at employers

Product track record

Skills

Now

Since April 2025 I have been independent: a self-directed technical deep-dive into applied AI while caring for a new child full time. Currently building PsychBill, billing automation for Canadian psychiatrists (pre-revenue, Phase 1 in development). Open to product leadership roles where AI capability is the product: B2B SaaS adding real AI to real workflows, founding product roles, or applied-AI consultancies.