Desk OS

Analysts in sanctions, fraud, regulatory, and operations queues work across several systems of record that each rank by their own logic. Holds get buried, incomplete cases bounce around without clear asks, and false positives never feed back into tuning.

A Next.js process overlay with a ranked shift board, per-desk policy lanes, case packets with missing-evidence asks, a deterministic Ask Desk panel, a disposition ledger, and a Learn view that turns overrides into retune candidates. Live demo with seeded banking, fraud, regulatory, and engineering queues.

Banks and operators already have four systems of record: core banking, transaction monitoring, GRC tools like IBM OpenPages, and manufacturing execution. Analysts do not need a fifth. They need to know which case to open first, what evidence is missing, and what they are not allowed to clear. Desk is a process overlay that reads exports from those systems, normalizes them into one ranked board, and runs a single loop: Identify, Prioritize, Act, Learn. It writes back labels only. It never moves money.

1. 1.
   One Ranked Board: Normalize cases from banking, fraud, regulatory, and engineering sources into a shift board ranked by look-now priority, with policy holds pinned above score.
2. 2.
   Severity Is Not Priority: Keep harm-if-true (severity) separate from look-now (priority). A low-severity case can still be first in the queue when a cutoff is near, and a severe one can wait if it is already parked with an owner.
3. 3.
   Need More as a Packet: When a case is incomplete, Desk does not just flag it. It builds a packet of specific asks, such as a second identifier on an OFAC near-match, so the next person knows exactly what to chase.
4. 4.
   Learn Without Auto-Clearing: Turn analyst dispositions into retune candidates for the source system. Repeated false positives cost less attention next time, but they still appear, and holds and elder-abuse cases can never be tuned off.

• ◆
  Policy Lanes in Data: Each desk has policy rules (sanctions hold, payroll window, wire cutoff, exam findings) that set priority floors and mark cases as never-auto-dismiss. Rules live in data, not scattered through UI code.
• ◆
  Adapters as Manifests: Every connector is a manifest that maps source fields to a case, lists the critical fields a human needs, and quotes evidence straight from the source row. A renamed column means editing the manifest, not the code.
• ◆
  Deterministic Ask Desk: The chat panel has no model behind it. It answers why a case is ranked first and what is missing, and it refuses requests like auto-releasing a held wire. Predictable beats clever in a sanctions queue.
• ◆
  Eval Before Demo: 117 automated checks cover deep intake, locked holds, Need more packets, deterministic chat, and park-with-owner, so the guarantees the demo makes are tested, not narrated.

Every new system of record asks a regulated team to migrate data, retrain staff, and re-validate controls. Most never get past procurement. An overlay reads what already exists, adds ranking and process discipline on top, and writes back nothing but labels. The test is simple: would an analyst walk into this board on Monday instead of their source queue? If they already know their top case, Desk is not for that desk.

Desk will not auto-release money, will not dismiss a P1 case without a ledger note, and will not let a parked case disappear from the queue. Parking means monitored with a named owner, not done. Marking it done would be a lie. Those refusals are the product, because the cost of a silent drop in a sanctions or exam queue is far higher than the cost of one more click.

The design started from what an analyst must never get wrong, not from features. Holds cannot auto-clear, parked is not done, and learning can downweight noise but never hide it. Everything else is ranking and packaging around those rules.

Live demo with four seeded desks and a 117-check eval suite covering holds, intake, Need more packets, and deterministic chat.