MatrixMint

MatrixMint turns RFPs into an evidence-anchored compliance matrix + bid-ready packet in minutes—powered by Gemini, with proof-verification, caching, and graceful fallbacks for judge-safe demos.

Comment

Inspiration

RFPs are expensive to respond to and easy to get wrong: teams waste days mapping requirements, and proposals still get rejected for missed details or unsupported claims. MatrixMint was inspired by a simple goal: make procurement responses faster and more defensible by forcing every claim to be tied to evidence.

What it does

MatrixMint ingests: • an RFP (requirements) • a Capability Brief (your “source of truth” evidence)

It produces: • a compliance matrix (Covered / Partial / Missing) with concise responses • evidence IDs + quotes per covered requirement • a proof score that verifies quoted evidence actually exists in the Capability Brief • bid-ready exports (proof pack, proposal draft, clarifications email, risks CSV)

It also supports deterministic “judge-safe” execution: • LIVE: Gemini analysis • CACHE: instant replay of a previous run • OFFLINE: deterministic fallback when quotas/timeouts happen

How we built it

• Next.js API routes handle orchestration and exports.
• Gemini API generates a strict JSON compliance result from the RFP + evidence-only Capability Brief.
• A proof verifier normalizes text and checks that each evidence quote appears in the Capability Brief (supports ellipsis/wildcard matching).
• A self-healing proof step repairs mismatched quotes by extracting deterministic snippets around the referenced evidence IDs—keeping outputs verifiable without inventing capabilities.
• A two-tier cache (memory + disk on /tmp in serverless) enables repeatable demos and rapid iteration.
• A quota/overload circuit breaker and live-call gate reduce flakiness during back-to-back judge runs.

Challenges we ran into

• Serverless filesystem constraints (read-only deploy FS) required moving persistence to /tmp and making caching robust across warm/cold instances.
• Model output variability: even with JSON schema constraints, strict parsing and retry logic were needed.
• Proof integrity: “sounds right” isn’t enough for procurement—evidence needed to be machine-verifiable, and failures needed safe, deterministic repairs.

Accomplishments that we're proud of

• Evidence-verifiable outputs: every “Covered” row includes evidence IDs and quotes that are automatically checked against the Capability Brief.
• Bid-ready artifacts generated directly from the analysis (proposal outline, clarification email, risk CSV, proof pack).
• Judge-safe reliability: caching + offline fallback allows consistent evaluation even under quota or transient model/API issues.

What we learned

• Reliability is a feature: in real procurement workflows, “works most of the time” is not acceptable.
• Adding verification (proof checks) changes behavior: the system becomes more conservative and more trustworthy.
• Serverless environments reward designs that degrade gracefully and keep evaluation paths deterministic.

What's next for MatrixMint

• Multi-document evidence packs (policies, SOC2 excerpts, prior case studies) with structured citations.
• Requirement clustering and duplicate detection across large RFPs.
• Collaboration workflows (review/approve, redlines, export to Word/PDF) for real proposal teams.

Built With

Share this project:

Updates