Chase Beck · Denver, Colorado
AI In Sensitve Spaces
I'm a senior engineer who ships AI into systems where being wrong can be expensive. Fifteen years across cannabis wholesale/compliance, and healthcare - regulated, payment-hostile markets where auditability and correctness are the hard part, not scale.
What I take on
Most AI engineering talent is Python-native. Most of the companies that need AI features are not - some run mature applications carrying a decade of domain logic, real customers, and nobody in-house who has shipped an LLM feature to production. The work lives in that gap.
It usually takes one of three shapes.
- An audit One week, fixed price
- I read your codebase and come back with the highest-leverage AI features for your product, what each costs to build and to run, what in your architecture blocks them, and a working proof of concept of the most promising one. If it is useful you keep going with me. If not, you keep the document.
- A feature, shipped Two to four weeks
- A scoped engagement that ends with something in production - including the evaluation harness that tells you it still works after a model version changes. That harness is the part most teams skip, and it is why their first AI feature quietly degrades six months later.
- Ongoing ownership Eight to ten hours a week
- I own the AI surface of your product: reviewing your team’s LLM code, keeping evaluations green as models shift underneath you, and clearing the upgrade debt that is blocking the work.
If your data is regulated - health, financial, anything with a compliance reviewer attached - that is the version of this problem I know best.
Where the work actually lives
The hard part of production AI is rarely the model call. It is the boundary: what crosses it, who owns it, and what happens when the system on the other side is slow, non-deterministic, or confidently wrong.
My answer keeps your backend as the system of record and the orchestrator. Inference lives in the ecosystem its libraries actually come from, rather than being forced into the monolith or bolted on as a rewrite. One engineer stays accountable for the seam between them, and for every client sitting in front of it.
In practice that means asynchronous dispatch to a slow upstream, a return path that never holds a web request open for five minutes, authorization enforced at the service boundary rather than assumed, and one API serving both a web application and a native mobile client.
Track record
Jihi
HIPAA-compliant telehealth and wellness platform serving licensed providers and thousands of users across web and mobile. Technical owner of the entire product surface - schema, API, background processing, real-time messaging, the Flutter client, and deploys.
Hotel Engine
Rails marketplace handling hundreds of bookings a day. Led the migration of stored cardholder data to Stripe and implemented Spreedly tokenization across a multi-vendor payment stack - zero downtime, touching every transaction in the business.
Helix TCS
Ran the acquired platform through integration into BioTrack and Canalytics, and wrote the documentation and test coverage the inheriting team needed to own it.
Cannabase
Wholesale cannabis exchange built in a newly legal, bank-hostile market. Grew to roughly 70% of Colorado’s licensed cannabis businesses and owned the market pricing dataset the trade press cited. Acquired by Helix TCS.
Get in touch
I'm taking on a limited amount of contract work. If you have an application making real money and an AI feature you have not been able to get into production - especially if your data is regulated - I would like to hear about it.