Lacuna

Women's health M&A diligence stack — verified deals, genomics governance, cited analytics

Curated dataset · n=58 verified deals · Not live market data · Scores are descriptive, not forecasts.

The live app reads src/data/dataset.verified.json by default. Server-side LLM calls use Vercel AI Gateway only. TensorFlow code is quarantined (not in the app). See MODEL_CARD.md before citing any score.

Table of Contents

• Overview
• What is Lacuna?
• Live Demo
• Core Features
• Descriptive analytics
• Health Equity context
• Clinical Trials
• Genomics variant store
• Academic frameworks
• Typography
• Technology Stack
• Quick Start
• Data Curation
• Documentation
• License

---

Overview

Lacuna is an investment research stack — a diligence infrastructure prototype with a curated, source-linked snapshot of women's health M&A (58 verified deals), rendered as D3 network views and descriptive analytics with published methodology.

Claim	Reality
Deal data	Static dataset.verified.json (manual verification from SEC, press, filings)
Scores & "predictors"	Deterministic rules and small-n statistics — MODEL_CARD.md
"ML" / TensorFlow	Quarantined under src/lib/ml/_quarantine/ — not imported by the app
Server LLM	INFERENCE.md — Vercel AI Gateway (+ OpenAI fallback for local dev)
Clinical trials panel	Live ClinicalTrials.gov search; M&A panels still use the curated dataset
Production intelligence	No — not PitchBook, not a data SLA, not investment advice

Open source under BSL 1.1 for corp VC diligence workflows, portfolio review, and self-hosted exploration. Commercial competitive products need a separate license — mps5cy@virginia.edu.

Portfolio project by Mae Kass.

Deployment: The analytics product runs on Vercel (this repo). A separate Framer site is for brand and narrative only, with one primary CTA into the live demo — see SITE_ARCHITECTURE.md and the framer/ build kit.

---

What is Lacuna?

An open-source diligence prototype for corporate VC and healthcare investors exploring verified women's health / FemTech M&A:

• D3.js force-directed acquirer–target graphs
• Deal flow & valuation charts (curated counts and disclosed values)
• Descriptive scoring (factor weights, cosine similarity, k-means — no trained forecast models in the UI)
• ClinicalTrials.gov lookup (live API; separate from deal JSON)
• Health-equity context with cited disparity statistics (descriptive, not allocation advice)

Every analytical panel in the app shows the provenance line above.

---

Live Demo

lacuna-maekass.vercel.app

Resource	Link
Application	lacuna-maekass.vercel.app
Repository	github.com/maekass/Lacuna
Methodology	docs/MODEL_CARD.md
License	BSL 1.1 → Apache 2.0 May 2030

---

Core Features

Verified deal explorer

• 58 verified acquisitions (fertility, oncology, diagnostics, menopause, pelvic health, precision medicine)
• Acquirers include Hologic, KKR, Pfizer, Gilead, Boston Scientific, and others named in sources
• Dataset v5 · updated per provenance.lastUpdated in JSON
• Sources: SEC EDGAR, press releases, investor relations (see DATA_CURATION_CHECKLIST.md)

Interactive network (ForceNetwork.tsx)

D3 force-directed graph: sector colors, deal-type edges, valuation-scaled nodes. Methodology: NETWORK_ANALYSIS_METHODOLOGY.md.

Deal flow (DealFlowChart.tsx)

Year-over-year counts from verified announcedDate — animated bars, no synthetic deal generator.

Valuation matrix (ValuationMatrix.tsx)

Sector × stage heatmap using disclosed values only; cells show company counts and averages.

Exit-likelihood leaderboard (QuantValuationPanel.tsx)

New: Heuristic valuation and exit-likelihood section with:

• ValuationEngine — bounded comparable multiples (EV/Revenue, EV/EBITDA) with uncertainty disclosures
• AcquisitionPredictor — sector-stage acquisition probability estimates (15/75 coverage noted)
• HealthImpactModeler — lives-saved modeling with Cohen's d bounds (not a rate)
• PortfolioOptimizer — stage-varying risk-adjusted ROI optimizer
• Verified-fields-only adapter — adaptQuantCompany uses only validated dataset fields; absent inputs remain undefined per provenance rules

See MODEL_CARD.md for methodology and caveats.

---

Descriptive analytics (heuristics, not predictive ML)

Curated dataset · n=58 verified deals · Not live market data · Scores are descriptive, not forecasts.

Acquisition likelihood indicators (ExitPredictor.tsx)

Transparent factor scoring for non-acquired companies in the verified set. Fixed weights, full disclosure in UI and MODEL_CARD.md. Not a predictive model; no TensorFlow.

Company similarity (CompanySimilarity.tsx)

8-D feature vectors, inline cosine similarity — "companies like this" for exploration.

Clustering (ClusteringAnalysis.tsx)

k-means on valuation × employees — descriptive segments (Emerging / Growth / Late-stage labels).

Optional server narratives (INFERENCE.md)

• UI blurbs via POST /api/ai/insights → Vercel AI Gateway (anthropic/claude-sonnet-4 slug).
• Exploratory copy only — heuristic scores on the curated dataset remain authoritative.

---

Health Equity & Black Women's Health

Descriptive context on disease areas with documented disparities and public market-size estimates — for learning, not buy/sell recommendations:

Disease	Disparity (cited in docs)	Public market-size estimate
Maternal Health	Higher mortality disparity	$12B
Uterine Fibroids	High prevalence	$34B
Lupus	Higher prevalence	$8B
Sickle Cell Disease	Population concentration	$5B
Cardiovascular Disease	Higher mortality	$15B

See OAIS_METHODOLOGY.md for scoring limits.

---

Clinical Trials Integration

• Live: /api/clinical-trials → ClinicalTrials.gov API v2 (search, batch lookup)
• Curated M&A: unchanged — still dataset.verified.json

Do not conflate live trial search volume with verified deal coverage.

---

Genomics variant store (optional)

Large VCF/gVCF call sets use a two-tier layout (off by default on Vercel):

Tier	Technology	Contents
Object storage	Local data/variants/ or S3	Multi-GB raw VCF blobs
Variant catalog	ClickHouse	Callset metadata + queryable variant summaries

• Dashboard: VariantCallsetBrowser — browse callsets, filter by gene, presigned S3 download when configured
• APIs: /api/genomics/callsets, /api/genomics/variants, /api/genomics/callsets/{id}/object
• Ingest: npm run clickhouse:ingest-vcf — stream parser → object storage → batch INSERT
• Docs: GENOMICS_VARIANT_STORE.md

docker compose up -d clickhouse
# .env.local: LACUNA_VARIANT_STORE=clickhouse, CLICKHOUSE_URL=http://lacuna:lacuna@localhost:8123
npm run clickhouse:migrate && npm run clickhouse:seed
npm run dev

Not clinical-grade genomics infrastructure — infrastructure demo with honest provenance labels.

---

Academic Frameworks

Six frameworks with explicit small-n limits documented in docs/ (causal DAG, fairness audit, network concentration, etc.). We state what cannot be claimed with n≈58 deals — see methodology files linked from the app.

---

Typography

The live app loads Playfair Display (Didone serif) via next/font/google and applies it app-wide — body copy, headings, and font-mono utilities share the same family for a high-contrast editorial look.

GitHub does not load custom web fonts. This README uses a Didone fallback stack (Didot, Bodoni MT, Georgia) so the page reads closer to the product on github.com. Only the live demo renders true Playfair Display.

---

Technology Stack

Layer	Used in production UI
Playfair Display (next/font/google)	App-wide Didone serif typography
Next.js 16, React 19, Tailwind v4	App shell
D3.js v7, Framer Motion	Visualization
simple-statistics	Descriptive stats / similarity / quant engine
Verified JSON (getVerifiedDataset())	Default data path; static import for Vercel serverless
PostgreSQL	Optional LACUNA_DATA_MODE=db
ClickHouse + S3/local object storage	Optional variant call-set catalog (LACUNA_VARIANT_STORE=clickhouse)
Vercel AI Gateway + AI SDK	Optional narratives + SEC classification (INFERENCE.md)
TensorFlow.js	Quarantined — devDependency for Vitest only
Deno (CI)	deno fmt and deno lint in GitHub Actions

CI Status: deno fmt, deno lint, eslint, vitest (297 tests), next build + tsc all green on main.

---

Quick Start

git clone https://github.com/maekass/Lacuna.git
cd Lacuna
npm install
npm run dev
npm run validate:dataset
npm run infra:check
npm test

Open http://localhost:3000. Data loads from src/data/dataset.verified.json unless LACUNA_DATA_MODE=db is set and Postgres is provisioned.

Optional local Postgres: docker compose up -d → copy .env.example to .env.local → npm run db:migrate && npm run db:import. See INFRASTRUCTURE.md.

Optional variant store: docker compose up -d clickhouse → set LACUNA_VARIANT_STORE=clickhouse → npm run clickhouse:migrate && npm run clickhouse:seed. See GENOMICS_VARIANT_STORE.md.

---

Data Curation

Manual verification — no synthetic maDeals. Workflow: DATA_CURATION_CHECKLIST.md, npm run validate:dataset, optional npm run sec:scan.

---

Documentation

Doc	Purpose
MODEL_CARD.md	Start here — what each score is and is not
INFERENCE.md	Server-side LLM (AI Gateway)
DATA_CURATION_CHECKLIST.md	Schema, validation, staging
NETWORK_ANALYSIS_METHODOLOGY.md	Graph metrics, small-n
OAIS_METHODOLOGY.md	Health impact scoring limits
INFRASTRUCTURE.md	CI, Vercel, Postgres, cron, /api/health
PERFORMANCE.md	Bundle, caching, probe split, fan-out limits
GENOMICS_VARIANT_STORE.md	ClickHouse + object storage for large VCF catalogs
MONITORING.md	Uptime URL: /api/health only (not /ready)
PRODUCTION_SETUP.md	Vercel env vars and migrations
SEC_INGESTION.md	SEC EDGAR cron pipeline
SITE_ARCHITECTURE.md	Vercel product vs Framer marketing (no analytics in Framer)
framer/BUILD_GUIDE.md	Framer marketing site — copy, tokens, HTML prototype
AGENTS.md	Contributor conventions

---

License

BSL 1.1 — research/education production use allowed; Competitive Offerings (commercial women's-health M&A intelligence products) require a separate agreement. Converts to Apache 2.0 May 2030.

mps5cy@virginia.edu for commercial licensing.

---

Author

Mae Kass — open investment-research tools for women's health data literacy and honest analytics.