Deal intelligence control plane 01 / 05
Your CRM says commit.
Does the buyer?
This system compares seller confidence with exact buyer language, activity, stakeholder coverage, and dated commitments. It then drafts a rescue plan that cannot act until a sales manager decides.
Live decision room
Test forecast against evidence
Ready
Awaiting evidence
Run a deal to reveal the forecast gap.
The same interface will show seller position, evidence position, exact buyer language, graph trace, and approval state.
Tracing buyer evidenceNormalizing source receipts
RUN BLOCKED
The analysis did not complete.
Check the request and try again.
Synthetic deal
Deal analysis
Forecast truth
Seller position vs buyer evidence
Priority risks
0 findingsBuyer objections
Exact languageHuman gate
Manager decision required
Review the evidence before recording a decision.
A decision receipt is stored. CRM, email, tasks, and calls remain untouched.
Golden scenarios12 / 12Expected risk and forecast
Evaluation protocol15 / 15All metrics at target
API and policy tests53 passed83.97% code coverage
External mutations0 allowedManager receipt only
System x-ray
Evidence first. Agents second. Action last.
Deterministic code owns privacy, source integrity, forecast policy, authorization, and action boundaries. LangGraph coordinates the bounded analysis roles around those controls.
01CRM + transcript + activitySigned synthetic receipts
02Data contractPrivacy, hashes, injection quarantine
03 · parallel analysis
ConversationObjectionsStakeholdersEngagementCommitments
04Forecast truthEvidence category + rescue plan
05Manager receiptNo system mutation
Browser and Postman friendly contracts, Pydantic validation, and generated OpenAPI without a separate schema layer.
Explicit parallel nodes, checkpoints, a visible approval interrupt, and deterministic state transitions that can be tested.
Tenant-scoped runs, immutable evidence references, objections, decisions, usage, and hash-chained audit records.
Open the architecture library4 rendered and editable views
Technical proof
Built as a deployable control plane
One command runs the API, PostgreSQL, durable LangGraph checkpoints, migrations, health checks, and public safeguards. The Vercel build uses signed stateless receipts for serverless approval continuity.
Inspect the OpenAPI contract
Python 3.12
FastAPI
LangGraph
PostgreSQL
OpenTelemetryEvaluation and observability
Every claim leaves a receipt.
Evaluation metrics
Risk detection, top-risk accuracy, forecast accuracy, schema conformance, evidence binding, injection resistance, policy accuracy, action safety, exposure bounds, and approval integrity.
Operational signal families
HTTP latency, run duration, active runs, node executions, findings, errors, quota rejection, policy results, approvals, checkpoint resumes, risk, exposure, and audit events.
Evidence claim classes
Observed, inferred, and recommendation. Unsupported certainty is disclosed. Instruction-like transcript text stays data and never becomes control.
Public demo boundary
Real system behavior. Synthetic business data.
- Included Typed APIs, graph execution, evidence receipts, evaluations, metrics, approvals, Docker, PostgreSQL, and CI.
- Not included Client records, buyer identities, CRM credentials, call recordings, outreach, task creation, or production write access.
- Pilot path Customer-owned deployment, read-only connectors first, explicit data approval, staging UAT, feature flags, cost ceiling, and rollback.