Obin
A verification workspace that turns an AI-drafted credit memo into claims, evidence, and decision points for an analyst.
- Year 2026
- Client Obin · concept
- Role Product design · Front-end prototype
- Role: Product design · Front-end prototype
- Client: Obin · concept
- Year: 2026
Project highlights
38
figures in one memo
28
proved without a person
2
drawn back at random, unnamed
I designed and built a concept workflow for reviewing an AI-written credit memo. It makes every machine assertion inspectable, orders unresolved claims by risk, and deliberately returns sampled claims to a person instead of presenting automation as certainty.
Agentic AI · FintechProduct design · Front-end prototype · Obin · concept
Obin is a concept workflow I designed and built around one question: after an agent writes a credit memo, how does an analyst know what deserves attention? Every machine-written figure carries its evidence state. The review queue starts with the highest-risk unresolved claim.
The workflow decision: draft fast, expose evidence immediately, then stop when judgment is required.
The interface makes authorship and evidence state legible at reading speed: human writing in serif, machine assertions in monospace, and an underline for proof, conflict, or missing support.
The review decision: risk order puts the most consequential unsupported claim first. Opening it reveals the source, the agent's reasoning, and what is still missing.
The working prototype connects four states that are usually split across tools: the agent run, the drafted memo, a risk-ordered verification queue, and the source record behind each figure. The analyst can move from an assertion to its evidence without leaving the memo, then advance, hold, or pass the review with private notes attached.
Obin is a designed and built concept, not a deployed lending product. It demonstrates a review model in which automation makes its evidence visible and stops at judgment. The next step is testing the queue with venture debt analysts to learn whether the risk ordering matches how they investigate a memo under time pressure.
Open any figure to see what supports it, what conflicts, and what still needs a person.
Back to top↑The thing I keep coming back to is that the agent made the analyst’s job worse before it made it better. Four hours became four minutes, and what came out the other side was a well written memo with thirty eight numbers in it and no way to tell which the machine had invented. So the analyst checked all thirty eight. The speed went somewhere, but not to them. The first test would examine the ordering assumption: whether debt load crossed with evidence reliability is the right sort. The workflow is a concept, so the interface identifies inferred behavior instead of presenting it as validated.
Next project Ledger→ A reviewer has twenty three submissions, one sitting, and a dozen browser tabs per cand... Back to top↑