Applied AI for payment operations

Runtime Lab

We build the harness payment teams run mission-critical work on, and study what makes agents reliable enough to touch money: evals, routing, failover, and memory, tested on real payment operations.

What we work on Field notes

What we work on

Frontier models are general. Payment operations are not. Our work is the layer between them that makes an agent safe to put on a live queue.

eval · merchant-re-reviewv7

#4821root cause found✓ pass

#4822evidence cited✓ pass

#4823stopped for approval✓ pass

#4824missing evidence✗ blocked

#4825reply matches policy✓ pass

#4826reserve within limits✓ pass

#4827sibling MIDs checked✓ pass

#4828sources linked✓ pass

Evals

Graded on real cases

Every skill is tested against resolved cases from your own queue. Did the agent find the root cause, cite the evidence, and stop at the approval point? A new model or prompt ships only if it passes.

router · evals decidelive

Routine enrichmentOpen 8B model

62% of runs$0.02 / run

Hard investigationsFrontier model

11% of runs$0.40 / run

−86%cost on routine work

Routing

The smallest model that passes

Each task is routed to the model and harness that still pass its evals. Routine enrichment moves to small open models, and hard investigations stay on frontier models.

failover drilllive

aws us-east-1

gcp us-west1

model primary

model fallback

09:41:02aws us-east-1 degraded

09:41:03failover → gcp us-west1

09:41:05run #8812 resumed

09:41:063 runs moved · 0 dropped

Reliability

Agents that stay up

We test recovery paths for sandbox and provider failures the way payment teams test disaster recovery, because a missed bank deadline costs real money.

memory · one paymentlive

Support

Payment ops

Finance

Compliance

MID 4410 · memory

✓Account changed Sep 18

✓Release approved

✓Sibling MIDs clean

Memory

Compounding organizational memory

Governed infrastructure that turns every investigation into memory the next case can use, with permissions, provenance, and an audit trail on everything an agent learns.

Visuals are illustrative.

Payments is a domain, not a prompt

Processor error codes, bank file formats, network reason codes, and sponsor bank quirks repeat across the industry. Our agents start out knowing them, so your team teaches them your business, not the basics. Your data, cases, and memory stay yours.

NACHA R01 · insufficient fundsNOC C01 · incorrect account numberVisa 10.4 · card-absent fraudMastercard 4837 · no cardholder authorizationISO 20022 pacs.008camt.053 bank statementBAI2 prior-day fileNACHA R10 · unauthorized debitFedwire IMAD/OMADRTP and FedNow returnsX9.37 image cash letterVisa 13.1 · merchandise not received

Field notes and research

What we learn putting agents on real payment queues, and which open models hold up on each job, written up so your team can try it.

Field notes

01What is an agent harness?02What is a system of intelligence?03How to reduce AI agent token costs 04How to reconcile payments with Claude Code and Codex 05How to investigate payment declines 06How to handle bank RFIs with agents 07How to automate merchant underwriting 08How to investigate fraud alerts 09How to investigate compliance alerts

Open model research

Best open source models for payment reconciliation Best open source models for partner-bank RFIs Best open source models for transaction monitoring Best open source models for chargebacks and disputes Best open source models for merchant underwriting Best open source models for KYC and KYB checks Best open source models for AML and compliance investigations Best open source models for payments customer support All open source models, ranked

Engineers who have built payment infrastructure

The team behind the lab built payments, risk, and reconciliation systems at fintechs and Fortune 500s, and works with customers as forward-deployed AI engineers.

Gus Trigos

Co-founder and CEO

AI engineer and second-time YC founder. Former quant in BlackRock’s AI Quantitative Investment Group. Built Mentum (YC S21, acquired) for Latin America’s largest asset and wealth managers: $20B+ in AUM data, 120k transactions a month.

Carlos Volante

Co-founder and CTO

Scaled financial infrastructure at Modern Treasury, Finix, and Hulu’s payments team (Disney): billions in volume, near-perfect uptime, and billing for tens of millions of users.

Yahoo

Berkeley

Backed by Y Combinator and fintech leaders San FranciscoRuntime on Y Combinator LinkedIn

Bring us your hardest queue

We’ll build the first agent with your team and measure it against the cases you already resolved.

Book a demo How we work with payment teams