Most AI quietly
stops working.
Systems that keep running in production.
Where it starts
We sit inside one process and map every step - what can be automated, what should not be, and what it costs you today.
Where it staysWe host, monitor and maintain what we built. Drift alerts, cost control, failure recovery, and a human on call.
The Keep
Nothing reaches your systems until it has cleared all five, and every gate writes down what it decided. Below: one morning at that desk.
The records
Not a diagram of a security layer, but the actual documents one run produced. This is what you would be reading at nine the next morning, and what an auditor would ask for.
Sample record. Anonymised, and representative of the format, not a real client’s claim.
req_a95b · claim 8841 · 3 August 2026 · 10 entries
Open any line for what was actually checked.
Now
Edit any line above. Watch both hashes.
claims-assessor · version 4 · in force since 12 June 2026
# Declared before the agent is allowed to run once. # Nothing here is decided at runtime, by the agent or by us. agent: claims-assessor version: 4 environment: your VPC · ap-south-1 · nothing leaves it may read claim records own business unit only policy library all payments ledger balances only may write claim records status, notes, recommendation may never payments ledger any write, any amount, any approver customer export any destination, including internal bulk delete any table outbound mail to addresses not on the approved list ceilings autonomous value ₹5,00,000 spend per run ₹5.00 retries 3 concurrent runs 20 escalates to role claims approver, on rota when value over ceiling confidence under 0.75 any two sources disagree # Changing any line above is a reviewed change with its own audit # entry. The agent cannot propose it, and we cannot make it alone.
req_a95b · raised 09:14:15 · signed 11:02:40
To: claims approver on rota · Re: claim 8841 · ₹42,00,000 · Raised: 09:14:15
I am not approving this. It is above my limit.
What I found
Policy CG-4471 is active and the loss is covered under §4.2. Invoice 8841 matches the surveyor’s report. There are three prior claims on this policy and none were disputed.
What I am not sure about
Two records disagree on the excess. The policy document says ₹50,000 and the CRM says ₹25,000. I used the policy document because it is the authoritative source, but I have not resolved which is correct and you may want to before you sign.
What I am recommending
Pay ₹42,00,000 less the excess. Confidence 0.91.
Everything above is in the audit record as req_a95b, including the reasoning behind each line. Approve, reject, or send it back for more work.
2026-07-19-002 · 2h 42m · 0 filings lost
Incident 2026-07-19-002 · customs filing gateway · detected by watch, automatically, eleven minutes before the client noticed
Impact: 14 filings queued · 0 lost · 0 duplicated · 0 partially submitted
11:40 gateway returns 502 on submit 11:40 attempt 1 failed, retry scheduled 11:41 attempt 2 failed 11:42 attempt 3 failed · retry ceiling reached, stopped 11:42 filings held in queue, none partially submitted 11:42 ticket raised, trade ops lead paged with the gateway’s own error 11:53 client informed, with an expected clearance time 14:20 gateway recovered, queue drained in original order 14:22 duplicate check run against customs reference numbers · clean
What changed afterwards
Gateway health is now checked before a batch starts rather than during it, and a queue deeper than ten pages immediately instead of at twenty-five. Both changes are in the manifest, with their own audit entries.
We publish these because the useful question is not whether something will break. It is whether you will hear about it from us or from your customer.
There are no client logos on this page. We are new, and we are not going to borrow anyone else’s.
Most of this industry sells you a demo and disappears at handover. We would rather show you the contract than a logo wall: fixed price, milestone payments, you own the code, and thirty days of stabilisation after we hand it over.
Founder, Surehand