MedReviewAI

AI-powered medical paper analysis. Compress scoping-review work from weeks to seconds.

Live demo: https://medai-deploy.vercel.app Status: Production-ready · College final-year project · Fully open data sources · Zero recurring API cost

Read the full audit report (PDF)

Comprehensive 30+ page report — UI screenshots, code excerpts, 10-persona testing log, accuracy benchmarks, full architecture. Hosted on the v1.0.0-audit release.

What it does

Upload a medical research paper (PDF / paste / search across 13 academic databases) → AI extracts:

PICO framework (Population, Intervention, Comparison, Outcome)
Demographics (sample size, age range, sex ratio, conditions)
Methodology (study design, duration, randomization, blinding, setting)
Outcomes (primary, secondary, statistics with p-values / CIs)
Confidence scoring (per-field 0-100 % calibrated to evidence strength)
Source-grounded references (every claim verified verbatim against the paper)

All results auto-save per user, exportable as JSON or CSV.

Demo URL

https://medai-deploy.vercel.app

Sign in with any email (Clerk dev mode, no email-verification required for testing).

Architecture

flowchart LR
  U[User Browser] -->|HTTPS| V[Vercel Edge]
  V -->|/api/*| F[FastAPI<br/>api/index.py<br/>Python 3.12 serverless]
  V -->|/* fallback| S[Static SPA<br/>React + Vite]

  F -->|verify_clerk_token| J[Clerk JWKS<br/>RS256 cached 1h]
  F -->|13 free APIs| EXT[External Sources]
  F -->|psycopg3| DB[(Neon Postgres<br/>analyses)]
  F -->|Llama 3.3 70B| GR[Groq API]

  subgraph EXT [13 academic sources]
    P1[PubMed]
    P2[PMC]
    P3[Europe PMC]
    P4[ClinicalTrials.gov]
    P5[PLOS]
    P6[OpenAIRE]
    P7[NIH RePORTER]
    P8[Semantic Scholar]
    P9[OpenAlex]
    P10[CrossRef]
    P11[DOAJ]
    P12[arXiv]
    P13[bioRxiv/medRxiv]
  end

  S -->|Clerk SDK| CL[Clerk Auth]
  S -->|Authorization Bearer JWT| F

Data flow — analyze a paper

sequenceDiagram
  participant U as User
  participant SPA as React SPA
  participant CL as Clerk
  participant API as FastAPI /api/analyze
  participant V as validate_grounding
  participant LLM as Groq Llama 3.3 70B
  participant DB as Neon Postgres

  U->>SPA: Click Analyze on paper
  SPA->>CL: getToken()
  CL-->>SPA: JWT
  SPA->>API: POST /api/analyze + Bearer JWT + (text | PDF)
  API->>CL: verify_clerk_token (JWKS)
  CL-->>API: sub claim (user_id)
  alt PDF input
    API->>API: extract_structured_text_from_pdf()<br/>section detect, dehyphenate
    API->>API: build_tagged_text() [SECTION: METHODS] etc
  end
  API->>LLM: prompt + tagged text (temp 0.15 JSON mode)
  LLM-->>API: structured JSON (PICO, demographics, ...)
  API->>V: verify each ref snippet against full text
  V-->>API: refs[].grounded + grounding_score
  API->>DB: INSERT into analyses (user_id = sub)
  DB-->>API: id
  API-->>SPA: JSON result
  SPA->>U: render PICO + Confidence + Source Refs

Screenshots

Live demo: https://medai-deploy.vercel.app

Hero

Animated Medical Papers gradient title, AI-Powered pill, theme toggle, dual primary CTAs (Upload a Paper / Search PubMed).

Features

Three core capabilities: PDF Analysis · PubMed Search · AI Extraction — refined glass cards with gradient icon badges.

Three Simple Steps

Input → AI Analysis → Results, connected by an animated dashed flow line.

CTA + Footer

"Ready to Analyze?" conversion card with shield badge + clean footer with quick links.

Full landing (single image)

Hero → Features → 3-step flow → CTA → Footer all in one shot.

Multi-source literature search (signed-in)

13 source chips · all free · zero API keys.

Search results — `metformin type 2 diabetes randomized controlled trial`

PubMed result cards with source badge, "open source ↗" link, abstract preview, Analyze button.

AI analysis — PICO framework

Real Groq Llama 3.3 70B extraction of a real PubMed RCT. Population, Intervention, Comparison, Outcome each in a coloured card.

Confidence + Source grounding

Per-field confidence (Population 90 %, Intervention 90 %), overall gauge, Source Grounding 100 % badge — every claim verified verbatim against the source text.

Full analysis page (single shot)

PICO → Demographics → Methodology → Outcome Measures → Confidence → Source References — all in one capture.

Results dashboard

History of every saved analysis, stat cards, JSON/CSV export, per-row delete.

Tech stack

Layer	Stack
Frontend	React 18 + TypeScript + Vite + Tailwind CSS + shadcn/ui
Animations	framer-motion + next-themes (light/dark)
Auth	Clerk (`@clerk/clerk-react` + `@clerk/themes`), JWT via JWKS
Backend	Python 3.12 FastAPI on Vercel serverless functions
LLM	Groq API — Llama 3.3 70B Versatile (temp 0.15, JSON mode)
PDF	PyMuPDF (fitz) — section-aware extraction
DB	Neon Postgres (psycopg3), per-user partition
Hosting	Vercel — single domain, 60 s function timeout

Features

1 · Search & Discovery — 13 free academic data sources

#	Source	Coverage	Auth
1	PubMed	NIH/NLM biomedical · 35 M+ citations	none
2	PMC Full-Text	Open-access medical full-text	none
3	Europe PMC	Broader OA biomedical + preprints	none
4	ClinicalTrials.gov	NIH registry · 480 k+ trial protocols	none
5	PLOS	Open-access journals (PLOS ONE / Medicine / Biology)	none
6	OpenAIRE	EU aggregator · 250 M+ research products	none
7	NIH RePORTER	NIH-funded project abstracts + investigators	none
8	Semantic Scholar	AI citation graph + TLDR summaries	none
9	OpenAlex	240 M+ scholarly works, all disciplines	none
10	CrossRef	DOI metadata + abstracts (130 M+)	none
11	DOAJ	Directory of Open Access Journals	none
12	arXiv	Preprints (incl. quantitative biology)	none
13	bioRxiv / medRxiv	Biology + medical preprints	none

Source-picker chips in the UI; each result card shows a source badge and an "open original" link.

2 · AI extraction (anti-hallucination)

Strict null-over-guess — model returns null instead of fabricating when evidence is missing
Verbatim-only source refs — every claim must include an exact substring quote from the source
Closed-enum study design — RCT / Cohort / Case-Control / Cross-sectional / Systematic Review / Meta-analysis / Qualitative / Case Report / Other
Closed-enum blinding — Single / Double / Triple / Open-label / Not applicable
Calibrated confidence — explicit 0.0–1.0 ranges with grounding-strength rules

3 · Source-grounding validation

Backend checks each LLM-produced quote against the source text (verbatim or ≥ 60 % word-overlap fallback)
Returns a per-paper grounding score (% of refs verified)
UI shows green "grounded" or amber "unverified" badge per ref

4 · Section-aware PDF pipeline

PyMuPDF text extraction → de-hyphenate line wraps → strip ligatures (ﬁ, ﬂ) → normalize whitespace
Regex-detect 18 standard medical-paper headers (Abstract, Methods, Results, Discussion, Limitations, etc.)
Build section-tagged input ([SECTION: METHODS], [SECTION: RESULTS], …) for the LLM, capped at ~14000 chars
Returns pdf_meta: { pages, sections_detected } for the UI

5 · Auth & security

Clerk JWT verified server-side via JWKS (RS256); cached 1 hour per Lambda instance
All 7 protected API endpoints reject without Authorization: Bearer → HTTP 401
Per-user data isolation via verified sub claim (used as DB partition key)
Old X-User-Id header path removed — clients cannot spoof identity
PDF size capped at 4 MB client-side (under Vercel's 4.5 MB body limit)

6 · UX

Light + dark mode (Clerk modal auto-syncs)
Custom logo (transparent PDF + ECG-pulse glyph)
Animated hero gradient, page transitions, loading screen, sleek pills
Fully responsive — phones / tablets / desktops
Document Viewer auto-flips horizontal/vertical at 768 px

Repo layout

medai-deploy/
├─ index.html                 # SPA shell, favicon links, anti-flash theme script
├─ src/
│  ├─ App.tsx                 # Routes, ClerkProvider, ThemeProvider, LoadingScreen
│  ├─ main.tsx                # ReactDOM bootstrap
│  ├─ index.css               # Tailwind base + theme tokens (light/dark) + utilities
│  ├─ lib/api.ts              # Typed API client + token-getter singleton + 13-source list
│  ├─ components/
│  │  ├─ ApiAuthBinder.tsx    # Registers Clerk.getToken at mount
│  │  ├─ DashboardLayout.tsx  # Sidebar + header for protected routes
│  │  ├─ ThemeToggle.tsx      # Sun/Moon switcher
│  │  ├─ LoadingScreen.tsx    # Boot-time splash
│  │  ├─ SourceReferences.tsx # Expandable refs with grounded/unverified badges + links
│  │  ├─ PicoTable.tsx, ConfidenceGauge.tsx, ExtractionCards.tsx, DetailedAnalysis.tsx
│  │  └─ ui/*                 # shadcn/ui primitives
│  └─ pages/
│     ├─ LandingPage.tsx
│     ├─ AnalyzerPage.tsx
│     ├─ SearchScreeningPage.tsx
│     ├─ ResultsDashboardPage.tsx
│     ├─ DocumentViewerPage.tsx
│     └─ AboutPage.tsx, NotFound.tsx
├─ api/
│  └─ index.py                # FastAPI app — all /api/* routes
├─ public/                    # favicon (SVG + multi-size PNG), apple-touch-icon
├─ pyproject.toml             # Python deps + requires-python pin
├─ vercel.json                # Rewrites + function config
├─ package.json               # Frontend deps + build scripts
├─ INTERNAL_AUDIT.md          # Full audit, metrics, security tests
├─ WHATSAPP_UPDATE.md         # Shareable team-update message
└─ README.md                  # This file

API endpoints

Method	Path	Auth	Purpose
GET	`/api/health`	public	uptime check
POST	`/api/search`	JWT	13-source academic search
POST	`/api/fetch-by-id`	JWT	Fetch single paper by PMID/DOI
POST	`/api/analyze`	JWT	PDF or text → Groq extraction → save
GET	`/api/analyses`	JWT	List user's analyses
GET	`/api/analyses/{id}`	JWT	Fetch specific analysis
DELETE	`/api/analyses/{id}`	JWT	Delete one
DELETE	`/api/analyses`	JWT	Clear all user's analyses

Database schema (Neon Postgres)

CREATE TABLE analyses (
  id              SERIAL PRIMARY KEY,
  user_id         TEXT NOT NULL,                    -- Clerk verified sub claim
  title           TEXT,
  year            TEXT,
  pico            JSONB,
  demographics    JSONB,
  methodology     JSONB,
  outcomes        JSONB,
  confidence      JSONB,
  source_refs     JSONB,
  input_type      TEXT,                             -- pdf | text | pubmed
  input_label     TEXT,                             -- filename | "PMID:12345" | …
  abstract_text   TEXT,
  summary         TEXT,
  key_findings    JSONB,
  clinical_significance TEXT,
  critical_appraisal    JSONB,
  takeaway_message      TEXT,
  analyzed_at     TIMESTAMP DEFAULT CURRENT_TIMESTAMP
);

All SELECT / DELETE statements filter by user_id = <verified JWT sub>.

Local development

Prerequisites

Node.js 18+
Python 3.12
A .env file at the repo root with the keys below

`.env` keys

VITE_CLERK_PUBLISHABLE_KEY=pk_test_xxx
CLERK_SECRET_KEY=sk_test_xxx
GROQ_API_KEY=gsk_xxx
DATABASE_URL=postgresql://user:pass@host/db?sslmode=require
PUBMED_API_KEY=optional

.env is gitignored. Never commit it.

Run frontend (Vite dev server)

npm install
npm run dev   # → http://localhost:8080

Run backend (local FastAPI)

cd backend
python -m venv venv
venv/Scripts/activate   # Windows
pip install -r requirements.txt
python main.py          # → http://localhost:8000

For local dev pointed at the local backend, set VITE_API_BASE=http://localhost:8000 in .env.

Deploy to Vercel

vercel deploy --prod

Python backend is auto-detected from api/index.py and pyproject.toml (Python 3.12 pinned).

QA & accuracy results

See INTERNAL_AUDIT.md for the full report.

Headline numbers:

13 / 13 sources operational, 130 papers retrievable per query
94.4 % field-level accuracy on synthetic RCT benchmark
100 % verbatim grounding on 3 separate authentic-paper tests
0 % hallucination observed
< 2 s PubMed search → AI analysis
< 5 s PDF analysis (11-page paper)
Sub-200 KB gzipped frontend bundle

Team

This is a college final-year group project. Contributors via GitHub:

License

Educational / academic project. Not for redistribution without permission.

Live URL: https://medai-deploy.vercel.app

Name		Name	Last commit message	Last commit date
Latest commit History 53 Commits
api		api
backend		backend
docs		docs
public		public
report		report
src		src
.env.example		.env.example
.gitignore		.gitignore
.python-version		.python-version
.vercelignore		.vercelignore
AUDIT.md		AUDIT.md
CONTRIBUTORS.md		CONTRIBUTORS.md
FINAL_REPORT.md		FINAL_REPORT.md
INTERNAL_AUDIT.md		INTERNAL_AUDIT.md
MedReviewAI_Report.pdf		MedReviewAI_Report.pdf
README.md		README.md
README_PRESENTATION.md		README_PRESENTATION.md
WHATSAPP_UPDATE.md		WHATSAPP_UPDATE.md
bun.lock		bun.lock
bun.lockb		bun.lockb
components.json		components.json
eslint.config.js		eslint.config.js
index.html		index.html
package-lock.json		package-lock.json
package.json		package.json
postcss.config.js		postcss.config.js
pyproject.toml		pyproject.toml
start_project.sh		start_project.sh
tailwind.config.ts		tailwind.config.ts
tsconfig.app.json		tsconfig.app.json
tsconfig.json		tsconfig.json
tsconfig.node.json		tsconfig.node.json
vercel.json		vercel.json
vite.config.ts		vite.config.ts
vitest.config.ts		vitest.config.ts

Folders and files

Latest commit

History

Repository files navigation

MedReviewAI

Read the full audit report (PDF)

What it does

Demo URL

Architecture

Data flow — analyze a paper

Screenshots

Hero

Features

Three Simple Steps

CTA + Footer

Full landing (single image)

Multi-source literature search (signed-in)

Search results — metformin type 2 diabetes randomized controlled trial

AI analysis — PICO framework

Confidence + Source grounding

Full analysis page (single shot)

Results dashboard

Tech stack

Features

1 · Search & Discovery — 13 free academic data sources

2 · AI extraction (anti-hallucination)

3 · Source-grounding validation

4 · Section-aware PDF pipeline

5 · Auth & security

6 · UX

Repo layout

API endpoints

Database schema (Neon Postgres)

Local development

Prerequisites

.env keys

Run frontend (Vite dev server)

Run backend (local FastAPI)

Deploy to Vercel

QA & accuracy results

Team

License

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Search results — `metformin type 2 diabetes randomized controlled trial`

`.env` keys