LoanWhiz

Our system directly answers the hackathon’s challenge – an open, SF-native agent framework with a library of domain primitives plus dynamic orchestration – by inverting the usual LLM pattern: it compiles the prospectus into a typed, executable deal model and lets deterministic primitives do every calculation, while the language model only orchestrates which primitives to run. The agent decides what to compute; it never does the arithmetic – so answers are reproducible and auditable, not hallucinated.

How complete it is:
This is a working end-to-end application, not a mock: a Docling + Gemini 2.5 Pro extraction pipeline; nine typed, governed primitives (prospectus extractor, ESMA tape normaliser, collections aggregator, waterfall runner, covenant monitor, report verifier, cashflow projector, audit logger, multi-period runner); a LangGraph orchestration agent; a Next.js analyst dashboard; and an MCP server exposing every primitive to other agents. It is validated to the cent against a real ING deal (Green Lion 2024-1 – 11/11 revenue steps, 4/4 redemption, €0.00 reconciliation gap), runs end-to-end across Dutch, Italian and Spanish RMBS, and we’ve included a live capability matrix to show exactly what’s proven and what isn’t – one deal validated to the cent, nine primitives exercised, and fifteen cells marked not-applicable, each with its reason – all backed by roughly 990 passing tests. Human-in-the-loop review is live, while the validator-driven auto-retry remains a wired hook rather than a closed loop.

Business model & long-term feasibility:
This wasn’t the challenge’s core so we didn’t spend too much time thinking about this. Having said that, the reusable-primitives-as-an-MCP design is the moat: deals are added as data, not code (drop JSON, no Python), so the same governed engine serves issuers, investors, trustees and rating analysts across jurisdictions. An open Apache-2.0 core drives adoption; a governed enterprise tier sits on top – FINOS-aligned model-risk controls and an evidence pack behind every answer – for regulated buyers who have to defend their numbers. And because it ships as an MCP server, it plugs into any host’s agent stack instead of living as a one-off script.

Technical excellence & pushing boundaries:
Every primitive returns a governed envelope (output + confidence + verbatim citations + append-only audit entry), and the registered primitive agrees with the platform to the cent – a direct MCP call matches the UI exactly. We integrate the organiser’s own tooling as load-bearing components, not name-drops: deeploans (Algoritmica’s ESMA ETL) as a selectable ingestion source and the FINOS AI Governance Framework for model-risk class, confidence gating and provenance – shipping a production-shaped, regulator-defensible prototype.

Just our in-house tools:

• liz (agent harness)
• deck-mcp (presentation-maker)
• prodemo-mcp (demo-maker)