Data rooms hold confidential deal documents. InsightDueDiligence is built to run inside a broker’s own environment, with the controls a deal practice, its IT team and its compliance function expect.
Runtime
Containers in your cloud
A FastAPI service and a static React UI, deployed in the broker’s own AWS account. Production schema changes go through Alembic migrations, then the engagement controls are seeded.
Data
PostgreSQL and object storage
Engagements in their own schema on the platform Postgres. Uploaded documents go to object storage under a key per workspace — never the API’s local disk. Database credentials come from the secrets manager by reference.
Models
Claude on Bedrock
Claude Haiku 5.5 for agent steps and Sonnet 5.5 for the copilot, in the broker’s own AWS account. Without a model the agents fall back to their checks’ findings and say so.
Identity, scope and audit
Sign-in through the Insight identity platform (OIDC), with roles and API keys. Service-to-service calls are signed.
Scope: a deal broker sees the engagements they own or are a member of; a practice leader, compliance or an administrator sees the practice. The copilot only reads engagements in the asker’s scope.
Every row belongs to an organization, and every change writes an audit entry: who, what and when, including what an agent did and as which agent.
Agents act as themselves, with their own identity, so the record always distinguishes an agent’s step from a person’s answer.
Model risk controls
1
Facts with sources
Each fact carries its page, benchmark or Pulse as-of date.
2
PII masked
Personal data is masked before anything reaches a prompt.
3
Budget & breaker
Per-agent daily budgets and a circuit breaker gate the call.
4AI
Claude
Answers in a fixed shape, citing the facts it used.
5
Valid and grounded?
Yes: the answer ships with its citations. Every number must be in the facts.
Over budget, or not grounded: the rules-only answer stands instead. Either way the trace records the model, tokens and cost.
Grounded figures. The engines compute every figure; a model answer with a number not in the facts is discarded.
Citations. Every finding cites a document page, a benchmark or an InsightPulse fact with its as-of date.
PII redaction before anything reaches a prompt; per-agent daily budgets; a circuit breaker on the model.
Traces and prompts. A trace per agent step with model, tokens and cost; prompts versioned with local defaults.
Autonomy capped. The ceiling ships at L2; every action that leaves the product is capped at ask first.
Tests never call a model. Bedrock is refused at the AWS client in the test suite, and each run uses a throwaway schema.
Documents
Uploads are checked for an allowed type, a matching file signature and a size limit before they are stored.
Without object storage configured, uploads are refused.
Documents can be marked privileged; data room search leaves them out unless asked explicitly.
Integrations
InsightPulse lookups are signed, time out, and are cached briefly; the key and secret are secret references.
InsightDealDesk hand-offs are an HMAC-SHA256 signed POST with a five-minute tolerance, idempotent per finding. With no endpoint or secret, hand-offs are held in the outbox and nothing leaves.
Outside production, actions that would send something (a request to the seller, a report, a submission) are dry runs.
Demo data
Everything shown on this site is an illustrative book: invented targets, carriers, people and figures, labelled illustrative in the product itself.
Talk to us about your environment
We will walk your IT and compliance teams through deployment, identity and the model controls.