A review-gated system that turns structured intake and evidence into a funding-focused business plan.
The challenge
Long-form business plans combine narrative, market judgment, financial assumptions, and lender-facing detail. A single prompt can sound polished while silently contradicting the underlying numbers, so the system needed explicit boundaries and release controls.
The system
- Validate structured intake and evidence before the first model call.
- Run market building and financial checking in parallel behind typed contracts.
- Compose the plan through a separately configurable writer step, including an Opus-quality model when the work calls for it.
- Critique, reconcile, optionally revise once, and preserve every draft, score, call record, and gate decision.
Important decisions
- Use typed records at every agent boundary so failures and retries remain inspectable.
- Keep financial arithmetic deterministic and ask models to exercise judgment around computed facts.
- Bound revisions instead of creating an open-ended agent loop.
- Block customer-final downloads until deterministic gates and human review conditions pass.
How the work moves
- Step 1
Validate intake
- Step 2
Build market and financial views in parallel
- Step 3
Compose the plan
- Step 4
Critique and reconcile
- Step 5
Review and release
Current evidence
- The internal release-candidate workflow runs end to end through CLI and web-demo paths.
- The system retains distinct audit artifacts, call telemetry, critic history, and rendered review packets.
- Commercial launch still depends on evaluator calibration, production OCR, and written SBA-counsel guidance.
What it taught me
- Good orchestration is mostly explicit contracts, deterministic gates, and honest failure states.
- A better writer model improves prose, but it cannot replace evidence quality or arithmetic controls.
- Auditability becomes a product feature when generated work informs consequential decisions.
Continue exploring
Return to selected work