AI & Automation · Agents & automation
Agents that check, scan, mark and search — and say "I don't know."
Audit routing on top of your existing robots, daily regulator scans on your own servers, marking against the official scheme, private search that abstains instead of guessing.
- 40% faster audits
- 4 h → 5 min
- 451 marking schemes
- 5 of 5 correct abstentions
What we build.
Audit and decision agents
Routing and decisions on top of your existing robots.
Case 11.
Regulatory compliance, on premise
A daily scan of regulator updates, on your own servers.
Case 17.
AI marking of handwritten papers
Marks against the official scheme.
Case 23.
Private search with a coverage gate
Says "I don't know" instead of guessing.
Case 29.
Research agents that grade sources
A cited research report from one question.
Case 28.
Video highlight finding
The moments worth clipping in a five-hour stream.
Case 30.
AI tutor inside an LMS
A study assistant that hands hard questions to a teacher.
Case 02.
How we keep it honest.
Grounded answers
Every reply cites your documents, and the agent abstains when it doesn't know.
Case 29: MRR 0.900, 5 of 5 correct abstentions.
Human handoff
A warm transfer to a person, and a do-not-call check before every dial.
Your keys, your data
You hold the API keys, and your data is never used for training.
On-premise option
Open models on your own GPUs, so documents never leave the building.
Cases 12, 17, 20.
Evaluation before go-live
Labelled test sets, automated evals, staged UAT, and approval gates on high-stakes actions.
Approval gates.
- 01Agent proposes
- 02Rule check
- 03Human approvalfor high-stakes actions
- 04Action
- 05Audit log
Payments, database writes and deletions never happen without an approval you set.
Agents we've built.
- Banking
Audit agents for 1,200+ bank users
40% faster audit cycles · HBL
- Compliance
Daily regulatory compliance, on premise
Daily scan of regulator updates · on-premise deployment
- EdTech AI
Marking handwritten exam papers with AI
451 marking-scheme PDFs registered · prototypes by Anees and the Zynthrix team · Mar to Apr 2026
- Private search
Search that says "I don't know"
1,169 video transcripts · lab · Sep 2026
- Research agents
Deep Search: a cited research report from one question
5 cooperating agents · lab, with CI · 2025 to 2026
- Video AI
Finding the moments worth clipping in a five-hour stream
4 signals fused · lab · Jul 2026
How it's priced.
- Sprint $490: one agent on one of your real documents or workflows.
- Build scoped per workflow.
- On-premise option.
Prices exclude taxes and third-party costs (hosting, carrier minutes, AI usage), billed at cost to your own accounts.
See all pricesQuestions about agents.
Will it hallucinate, and what happens when it does?
It can, so we build for it. Answers come from your own documents and cite them. When the agent can't find support, it says it doesn't know and hands over to a person: our private search abstained correctly 5 times out of 5 in testing. Payments, database writes and deletions always wait for an approval you set.
Who is liable when it's wrong?
Your contract sets liability, under English law. We also design so mistakes stay small: the agent says it's an AI, hands over when it's unsure, and can't make a payment, write to a database or delete anything without your approval.
Does it replace staff?
It takes the repetitive part: first replies, data entry, routine checks. Your staff handle what it hands over, and they sign off each part before it goes live.
How do you test it before it goes live?
On labelled test sets built from your own data, with automated evaluations, then staged UAT with the people who will use it. High-stakes actions stay behind approval gates after launch.
Send us one job.
Tell us what you need built. You'll have scope, price and dates in writing today, the work in 3–5 business days, and you pay only if you approve it.
Prices exclude taxes and third-party costs (hosting, carrier minutes, AI usage), billed at cost to your own accounts.