How it works

The run, the engines, the controls and the record — as the code does them

InsightDealDesk is a FastAPI service and a React workbench on the Insight platform. This page walks through the architecture, an Analyze run, the stage machine and what is shared with the rest of the suite.

Architecture

One product, built on shared platform packages

Who calls it

Workbench UIReact on the Insight UI kits

InsightDueDiligenceReviewed findings, as a signed hand-off

The service

InsightDealDesk APIFastAPI: pipeline, workspace, Analyze, stage machine, data room, quotes, binders

Shared platform packages

insight-brokerMarkets, quotes, binders, comparison, checking

insight-diligenceData room, playbook, findings, reports

insight-agentsPipeline, autonomy, registry, traces

insight-rulesRules Studio

insight-decisionsThe decision record

insight-assistThe deal copilot

insight-identitySign-in, roles, audit

Data and models

PostgreSQLSchema deal

Claude on BedrockIn your own AWS account, called by the agents

InsightPulseCompany intelligence, read by the diligence package

InsightXtractDocuments read into cited pieces

One service on shared packages: the same broker domain, agents, rules and record as the rest of the suite.

Domain

Broker domain

Roster and authority, clients, deals, markets and appetite, submissions, quotes, binders and fees; the comparison engine, binder checking and market matching.

Diligence

The data room and playbook

Documents read into cited pieces; the rw_placement playbook — 26 questions in six sections; findings with review; reports.

Agents

Pipeline and autonomy

Each agent runs as itself at its level; the autonomy service caps the ceiling at L2; every step traced in Lense.

An Analyze run

Eleven steps, each gather → reason → prepare

POST analyze

The broker presses Analyze; the API starts the pipeline’s eleven steps in order.

Diligence items

Data room, reps, gaps, exclusions and intelligence, each with its cited facts.

The engines

Market fit, quote comparison and the binder check: figures with their reasons.

4AI

Claude reasons

Over each item’s facts and the engines’ figures. Ungrounded answers are discarded.

Supervisor

Runs the placement controls; the run, its hits and the verdict go on the decision record.

What waits on you

Steps, the narrative and the prepared decisions, back in the workspace.

What would they do? runs the same pipeline inside a transaction that is rolled back: the broker sees what would be prepared, and nothing is written.

The stage machine

A deal moves forward only when its blockers are cleared

stateDiagram-v2
  [*] --> Engaged
  Engaged --> DataRoom: a workspace exists
  DataRoom --> Submission: SPA or APA in the room
  Submission --> Quotes: submission sent
  Quotes --> UWCall: quotes received
  UWCall --> Bind: no blocking control
  Bind --> PostClose: binder checked, sign-off done
  PostClose --> [*]

What the machine checks

Data room and later need a diligence workspace
Submission needs a purchase agreement in the data room
Quotes needs a sent submission
UW call needs a quote
Bind needs no blocking placement control
Post-close needs a practice leader’s sign-off where a control referred

A refused move lists every blocker. Moving back a stage needs a reason. Every move is audited.

Prepared decisions

One live ask per agent and action

stateDiagram-v2
  [*] --> Open
  Open --> Accepted: accept, carried out as the broker
  Open --> Changed: change with a note
  Open --> Rejected: reject
  Open --> Asked: ask a question
  Open --> Superseded: a newer run prepares the same action

The API

Everything the workbench does, the API does

Route	What it covers
`/api/deals`	Pipeline, workspace, analyze, prepared answers, stage machine, data room, search, findings review, reports, quotes, binders, audit
`/api/deal_agents`	Catalogue, runs and traces, autonomy, metrics, activity
`/api/deal_rules`	The Rules Studio: placement controls
`/api/deal_markets`	Markets and appetite
`/api/deal_intake/diligence-findings`	The signed InsightDueDiligence hand-off, contract `diligence_findings.v1`
`/api/home`, `/api/deal_meta`	Persona home; vocabularies

Models

Claude Haiku 5.5 for agent steps and Claude Sonnet 5.5 for the copilot, on Bedrock in your own AWS account. Without a model, agents fall back to their rule-based findings and say so.

Tests

Each run lives in its own throwaway schema. No test reaches a real model: Bedrock is refused at the client and counted; reasoning is tested with an injected transport.