Product

Roadmap

This prototype is framed as v1 of a real product. Below is what shipped, what's immediately next, what depends on learnings from Next, and what's explicitly out of scope.

Now

Shipped — This Prototype

Live

Live in this build.

Single-borrower diligence and monitoring workflow (Meridian SaaS Co., synthetic)
Three-agent pipeline: extraction → deterministic covenant monitoring → source-gated memo drafting
CEL-style policy gates enforcing calculated/interpreted separation and claim-to-source tracing
Eval harness with independently-sourced ground truth across extraction, breach detection, and memo hallucination
Live deployment: Railway backend + Lovable frontend, wired end to end

Immediate Post-v1 Priorities

Planned

Near-term, well-scoped work.

Multi-borrower portfolio view
Extend from 1 fictional borrower to a portfolio dashboard, surfacing cross-borrower risk concentration (e.g., "3 borrowers approaching runway covenant this quarter").
Human-review queue with SLAs
Convert the current UI flags into an actual assignable workflow for credit officers, with response-time tracking, not just a visual badge.
Expanded eval coverage at scale
The current breach detection eval is n=1 by design (single synthetic breach); Next means backtesting against a larger set of historical, anonymized real-world covenant scenarios to get a statistically meaningful precision/recall baseline.
Document format expansion
Real loan agreements are messier than the current clean markdown source; Next means testing extraction against scanned PDFs, inconsistent formatting, and amendments/side letters.

Later

Dependent on Next's Learnings

Horizon

Longer horizon, contingent on evidence.

Origination workflow integration
Extending beyond diligence/monitoring into deal sourcing and initial underwriting, closing the full lifecycle named in the original problem statement.
Cross-covenant correlation modeling
Detecting leading indicators (e.g., does a churn uptick reliably precede a burn multiple breach by 1–2 months across the portfolio?) rather than only reactive threshold monitoring.
Analyst feedback loop into the eval harness
When a human reviewer overrides an LLM interpretation, feed that correction back into future ground-truth refinement, closing the loop between production usage and eval quality — the flywheel described in the Frame-Prove-Earn model this build was designed around.

Scope

Explicit Non-Goals

Deliberate boundaries, not omissions

Full autonomous covenant enforcement (auto-triggering lender actions) — human approval remains structurally required at every stage where capital or legal rights are affected.
Replacing credit officer judgment on interpretive/ambiguous clauses — the system is designed to route these, not resolve them.
General-purpose document AI — this stays scoped to venture debt monitoring, not a horizontal contract-analysis platform.

Synthetic borrower data, portfolio demonstration only.