I cut a client's AI bill from $100K to $8K a month. Same product, same traffic, same quality bar. The spend was in the routing, the retries, and the models nobody had questioned.
That's the work: AI spend and assurance for B2B SaaS with AI in production. Your feature shipped. Now the bill has five figures in it that nobody can explain, and one day a customer, auditor, or board member will ask you to prove the outputs are correct. I make both questions survivable.
One week inside your production AI stack. $9,500, fixed fee. Five deliverables, in writing:
| 🗺️ Cost map | Where every dollar of your AI bill goes — by model, route, and feature |
| ✂️ Savings plan | Itemised changes your team can execute in normal sprints |
| ✅ Eval coverage report | What's tested vs. what's hoped |
| What breaks when providers change things | |
| 🔥 Failure mode inventory | How your system fails, and who finds out |
The guarantee: I identify at least 30% of your AI infrastructure spend in itemised annual savings — or the audit is free. No asterisks beyond that.
The fee is credited in full toward anything we do after. How it works →
The provider quietly swaps the model behind your endpoint. The bill creeps from $3K to $30K in increments that each looked small. Someone senior asks you to prove an output was correct on a specific date. None of these throw an exception — your monitoring watches for crashes, and AI doesn't crash. It rots.
I did tax compliance at PwC before seven years of software, the last three on production LLM systems. I've been on the other side of the table when an output is wrong and someone with authority asks why. Engineer enough to fix your system. Compliance enough to defend it.
- 💸 One client's monthly AI bill: $100K → $8K. Same product, same traffic.
- ✂️ Cut projected AI infrastructure costs by 35% through model routing analysis
- 🔬 Built the evaluation framework that verified output quality before a single user saw it
- 👁️ Shipped the observability layer that catches AI quality drift before customers do
- 🚀 Took a company's first generative AI feature from zero to production in 10 weeks
- ⏱️ Automated a manual workflow that was costing an ops team ~10 hours a week
You keep the written plan either way — run it yourselves, or keep me on. Most clients continue into a fixed-scope build of the monitoring and eval stack, or an ongoing assurance retainer: dashboards reviewed weekly, evals signed off before ships, the next model deprecation handled before it's a fire.
PwC (tax compliance) ──► Full-stack engineer ──► Senior SWE @ logistics SaaS ──► AI spend & assurance
2017–2018 2019 onward Golang/Node microservices, audits, evals, cost
3rd-party integrations, control — @animanova
testing strategy rebuild
If there's a line on your AI bill nobody can explain — or you couldn't prove your outputs were correct last Tuesday — start with the audit. Twenty minutes to scope it; I'll tell you what I'd expect to find.
⚡ Fun fact: I did tax compliance before I wrote software. Auditors made me the engineer I am — I build systems that answer questions before they're asked.