TLDR
Helped an AI video company go from stuck to shipped. They had an overcomplicated architecture and limited AI experience on the team. I simplified it, got something working on day one, and taught them how to systematically improve their agent.
The product later reached six-figure revenue.
The Problem
A company building an AI video agent came to me because they were stuck.
Their engineers had:
- Picked an immature framework that kept breaking
- Designed an overly complex agent architecture
- No prior AI/LLM experience
Iterations were painfully slow. Each change took a full day to test, and the team had been working for weeks without a usable product.
The Approach
Simplify first.
I reviewed their existing setup and proposed a much simpler architecture. Not because simple is always better, but because they were solving a straightforward problem with an unnecessarily complex solution.
The goal: get something working fast, then iterate.
What I Did
Day 1: MVP
Got something working in a single day. It wasn’t perfect, but the team finally had a usable baseline.
Week 1: Live Tuning
This changed how the team thought about iteration. I did real-time pairing sessions with the CEO and CPO:
- They’d chat with the agent
- Tell me what’s wrong
- I’d fix it immediately
- They’d try again
Iterations in minutes instead of days.
Their reaction: “We’ve been telling our engineers and iterations take one day. Here you are iterating in minutes. This is crazy. This is what we wanted.”
Month 1-3: Evals & Systems
Once the basics worked, I helped them build systems to improve on their own:
-
Web search tool evals: They didn’t know which search API to use (Tavily? Perplexity? Google?). Built simple evals, tested across tools, picked based on data instead of vibes.
-
Product-level evals: Set up evaluation pipelines for their video generation. How do you know if a generated video is good? How do you trace issues and improve? Taught them the systematic approach.
Results
- Something working in 1 day (vs weeks of being stuck)
- Iteration time: minutes instead of days
- Shipped to real users within the 3 month engagement
- Eval systems set up for continuous improvement
- Commercial outcome: the product later reached six-figure revenue
Without external help, getting a basic version would’ve taken 1+ month.
What I Learned
Complex frameworks feel productive but often just slow you down. Start simple, add complexity only when needed.
Async reviews are fine, but real-time iteration with stakeholders builds trust and ships faster. The CEO feeling the speed of iteration changed their whole mindset about what’s possible.
Related
- AI Consulting: More about my consulting work