Assistant + eval set
Assistants over your own documentation
Manuals, contracts, tickets and the ERP. Per-user permissions, a cited source on every answer, and an evaluation set that stays in your repository.
Your outsourced tech team · applied AI since 2009
AI systems
that make it to
production.
Sixteen years building software for companies, the last seven of them building AI that runs every day: assistants with real permissions, document automation, and models over your own data. Measured, not demoed.
- 2009
- First project shipped
- 1000+
- Projects delivered
- 100+
- Clients worldwide
- 1 in 4
- Audits that end with a no
Services
Six ways to start, every one of them measurable.
Flow in production
Document automation
Classify, extract and route invoices, delivery notes and policies. Verified against your own systems before anything gets written.
Model + pipeline
Models on your data
Hybrid retrieval, reranking, and fine-tuning when it earns its place. Reproducible training and versioned data from day one.
Deployed service
Computer vision
In-line quality control, reading difficult documents, counting. We start with labeling, because that's where the problem lives.
Gateway + platform
Infrastructure and cost
Your own model gateway, context caching, usage logging and spend alerts. The same layer that later carries your compliance story.
Report + sessions
Audit and training
Where AI fits in your operation, where it doesn't, and what each option costs. Includes the list of things we recommend not building.
How we work
Three phases. Each one decides whether there’s a next.
- 012 weeks
Diagnostic
We look at the actual data, systems and processes. You leave with use cases ranked by value and risk, and an estimate that holds up.
- 026 weeks
Measured pilot
One case, real users, and an acceptance threshold agreed before we start. If it doesn't clear the bar, we stop here and say so.
- 03ongoing
Production and handover
Deployment, observability, cost control and training for your team. The goal is that you can run it without us.
Work
Four projects, four industries.
Health insurer
1,100 employees · claims operations
38,000 claims a month classified and routed by hand.
- 64%
- routed with no human touch
- 9 → 2 days
- average resolution time
Industrial distributor
480,000 part catalog
Internal search couldn't find parts by standard or by equivalence.
- +31%
- conversion from search
- −46%
- calls to the technical desk
Corporate law firm
210 attorneys · vendor contracts
First-pass contract review, repetitive and expensive.
- 340 hrs
- of review saved per month
- 100%
- of findings with a verifiable citation
Consumer lending platform
Underwriting operations
They asked us to build a model that scores credit applications.
- −52%
- applications returned for paperwork
- 0
- automated credit decisions
Field notes
What we were learning, when we learned it.
AI on legacy systems: the work nobody demos
The model is the easy part. The hard part is getting the data out of an AS/400 with permissions, without breaking the monthly close.
Prompt debt: the code nobody dares to touch
Nine hundred lines of instructions, grown by patches, with rules that contradict each other. It's a recognizable pattern now.
Hybrid retrieval: BM25 never left
Part numbers, standards codes and proper nouns are still found better by keyword search.
Contact
One address. Write to it.
hello@pixel-myth.comWhat to send
The process that hurts, what systems hold the data, and the deadline you’re working to. Two paragraphs is plenty.
What happens next
A reply within one business day, and a call with no pitch in it. If AI isn’t the right tool for your case, we’ll say so on that call. See the whole flow.
Monthly note