Good funds get trapped in review. ~6 minutes per case becomes ~1 minute.
Same reviewer. Same final decision. The packet does the context work, the human approves — and everything uncertain is flagged, not hidden.
One held payout, one packet.
- 1Read your queue
Read-only on KYB, ledger, disputes, sponsor bank, policy, case system.
- 2Assemble the packet
Evidence, what's missing, similar cases, a draft recommendation.
- 3Reviewer decides
Your team approves, edits, or overrides — we never write to production.
- 4Corrections compound
Every edit feeds a queue-specific eval set, so the next packet sharpens.
Acme Goods
On top of your stack. Not in place of it.
We don’t replace Sardine, Unit21, Alloy, Middesk, Chainalysis, or your case system. They become evidence sources. We prepare the packet, and the human posts the approved update back.
payout hold time · backlog age · false-positive holds
evidence completeness · win rate · time to file
reserve coverage · release schedule · sponsor reporting
aged exceptions · failed-payment SLA · customer impact
manual-review rate · time to approve · missing docs
request turnaround · evidence completeness
For your engineering team →
Webhook or polling into your case system, JSON packet out. Pilot reads from a read-replica or a one-shot CSV; production adds a webhook back to your reviewer queue.
Packet ready in <60s p95 from case open. Cached evidence cuts repeat-case latency to <5s.
Every claim links to the underlying record id (KYB attestation, dispute, ledger entry, similar case). Audit any line in two clicks; replay any decision.
Frontier LLMs with a queue-specific eval harness. We pin models per queue and run regression evals on every prompt change. Reviewer corrections compound into the next packet.
Tell me about your queue.
Pilot shape, timing, and price are scoped per queue on a 20-min call — not pre-packaged. Read-only on your historical cases first; nothing touches production until you say so.
