LabProcure

From Hypothesis to Runnable Experiment with protocol design and lab procurement via Voice and Chat

Built for Hack-Nation 2026

What It Does

Paste a scientific hypothesis. In under 60 seconds, LabProcure returns a complete, operationally grounded lab experiment plan — with real catalog numbers, USD cost estimates, and a phased timeline. You can also talk to it via voice call to ask questions about the literature or the generated protocol in real time.

Three stages:

Literature QC — Semantic Scholar paper search + OpenAI novelty classification. Returns a novelty signal (not found / similar work exists / exact match found), up to 8 papers with abstracts, and reference URLs.
Experiment Plan — OpenAI with structured output generates:
- Step-by-step protocol (8–15 steps)
- Materials list with realistic catalog numbers (Sigma-Aldrich, ThermoFisher, etc.)
- Itemised budget with USD cost estimates (live Tavily price lookups)
- Phased timeline
- Validation criteria with quantitative success/failure thresholds
Feedback Loop — Every plan section is inline-editable. Corrections are embedded and stored in Supabase. The next similar experiment plan automatically reflects those expert corrections via few-shot prompt injection — no retraining required.
Voice Assistant(LabAgent) — A floating mic button on the Literature QC and Plan pages opens a live voice call powered by ElevenLabs Conversational AI. The agent knows your current papers and protocol, and can call search_literature (a client-side tool) to look up new papers mid-conversation.
Material Comparison & Email — Compare reagent alternatives via Tavily + GPT, and send supplier RFQ emails via Resend — all inside the Plan dashboard.

Tech Stack

Layer	Technology
Backend	FastAPI + Python 3.13
LLM	OpenAI via LangChain + OpenAI
Literature Search	Semantic Scholar API + Tavily fallback
Price Lookups	Tavily web search + LLM extraction
Embeddings	OpenAI `text-embedding-3-small`
Feedback Store	Supabase (PostgreSQL + pgvector)
Voice AI	ElevenLabs Conversational AI (WebSocket, client tools)
Email	Resend API
Frontend	React 18 + Vite + Tailwind CSS
Charts	Recharts
Animations	Framer Motion

Project Structure

ai-scientist-hacknation/
├── backend/
│   ├── main.py                    # FastAPI app + CORS + router registration
│   ├── routers/
│   │   ├── literature_qc.py       # POST /literature-qc
│   │   ├── generate_plan.py       # POST /generate-plan  +  /generate-plan/stream (SSE)
│   │   ├── chat.py                # POST /chat  — streaming text chat
│   │   ├── voice.py               # POST /voice-session  — ElevenLabs signed URL
│   │   ├── feedback.py            # POST /feedback
│   │   ├── compare.py             # POST /compare-material
│   │   └── email_quote.py         # POST /email-quote
│   ├── services/
│   │   ├── search.py              # Semantic Scholar + Tavily search + novelty classifier
│   │   ├── planner.py             # Staged plan generation + correction injection
│   │   ├── feedback.py            # Supabase store + vector similarity retrieval
│   │   ├── prices.py              # Material price cache + Tavily live lookup
│   │   ├── compare.py             # find_alternatives + get_substance_details
│   │   └── email_agent.py         # OpenAI tool-call RFQ composer + Resend sender
│   ├── models/
│   │   └── schemas.py             # All Pydantic models
│   ├── requirements.txt
│   └── .env.example
├── frontend/
│   ├── src/
│   │   ├── pages/
│   │   │   ├── InputPage.jsx      # Hypothesis entry screen
│   │   │   ├── QCPage.jsx         # Literature QC results + chat + voice
│   │   │   └── PlanPage.jsx       # Plan dashboard — protocol, budget, materials, chat, voice
│   │   └── components/
│   │       ├── VoiceCall.jsx      # Floating voice call button (ElevenLabs SDK)
│   │       ├── PlanChat.jsx       # Floating text chat panel for the plan
│   │       ├── InlineEdit.jsx     # Hover-to-edit with feedback POST
│   │       ├── ProductComparison.jsx  # Material alternatives UI
│   │       ├── GuideTooltip.jsx   # Onboarding tooltips
│   │       └── Toast.jsx          # Save confirmation notification
│   ├── vite.config.js             # Dev proxy → backend :8000
│   └── package.json
├── supabase_setup.sql             # Run once in Supabase SQL Editor
├── Justfile                       # Dev runner (just dev / backend / frontend)
└── README.md

Setup

1. Prerequisites

Python 3.11+
Node.js 18+
just command runner (brew install just)
API keys: OpenAI, Semantic Scholar, Tavily, ElevenLabs
A free Supabase project (for the feedback loop)
A Resend account (for supplier email)

2. Clone and install

git clone <repo-url>
cd mvp
just install     # creates backend/.venv + pip install + npm install

3. Configure environment

just setup       # creates backend/.env from .env.example

Edit backend/.env:

OPENAI_API_KEY=sk-...
SEMANTIC_SCHOLAR_API_KEY=...
TAVILY_API_KEY=tvly-...

# ElevenLabs — voice assistant
ELEVENLABS_API_KEY=sk_...

# Supabase — feedback loop + price cache
SUPABASE_URL=https://xxxxxxxxxxxx.supabase.co
SUPABASE_SERVICE_KEY=eyJ...

# Resend — supplier RFQ emails
RESEND_API_KEY=re_...
FROM_EMAIL=onboarding@resend.dev
SUPPLIER_EMAIL=you@example.com

4. Set up Supabase (feedback loop only)

Create a project at supabase.com
Go to SQL Editor → paste the full contents of supabase_setup.sql → Run
Copy the service_role key into SUPABASE_SERVICE_KEY in your .env

Disable RLS on the feedback table:

ALTER TABLE feedback DISABLE ROW LEVEL SECURITY;

5. Run

just dev         # starts both backend (port 8000) and frontend (port 3000)

Or separately:

just backend     # FastAPI on http://localhost:8000
just frontend    # Vite on http://localhost:3000

API docs: http://localhost:8000/docs

API Endpoints

`POST /literature-qc`

{ "question": "Does CRISPR-Cas9 efficiently edit plant cell genomes?" }

Returns novelty_signal, papers, references, context_summary.

`POST /generate-plan/stream`

SSE stream — emits staged progress events then a final plan JSON object.

{ "question": "...", "literature_context": "...", "references": ["..."] }

`POST /chat`

Streaming plain-text chat grounded in the current plan or literature context.

{ "messages": [...], "context": { "question": "...", "papers": [...] } }

`POST /voice-session`

Returns a signed ElevenLabs WebSocket URL for a live voice call session.

{}

Response: { "signed_url": "wss://...", "agent_id": "..." }

`POST /feedback`

{
  "experiment_question": "...",
  "category": "protocol",
  "item_label": "Step 3",
  "original_text": "Use 0.5% LAP photoinitiator",
  "corrected_text": "Use 0.1% LAP photoinitiator",
  "comment": "0.5% causes cell death in GelMA bioinks"
}

`POST /compare-material`

{ "material_name": "GelMA", "context": "bioprinting scaffold" }

Returns alternative suppliers and substance detail.

`POST /email-quote`

Composes and sends a supplier RFQ email via Resend.

Voice Assistant — How It Works

User clicks mic button
    └─ Frontend calls POST /voice-session
    └─ Backend creates ElevenLabs agent (cached) + returns signed WebSocket URL
    └─ Frontend connects via @elevenlabs/react SDK
    └─ onConnect fires → sendContextualUpdate injects current research context
    └─ Agent speaks, researcher replies

When agent needs more papers:
    └─ Agent emits client tool call: search_literature({ query })
    └─ Browser executes the tool → calls POST /literature-qc
    └─ Result returned to ElevenLabs → agent incorporates into answer

No server-side webhook needed — tool execution happens entirely in the browser via ElevenLabs' client-tool protocol.

Feedback Loop — How It Works

CORRECTION SAVED
Researcher edits Step 3: "0.5% LAP" → "0.1% LAP"
    └─ OpenAI embeds the experiment question → 1536-dim vector
    └─ Stored in Supabase: original, corrected, comment, embedding

NEXT SIMILAR EXPERIMENT
New question arrives → /generate-plan/stream
    └─ Embed new question → cosine similarity search (threshold: 0.70)
    └─ Match found → inject into GPT-4o system prompt:
       "Expert correction: use 0.1% LAP, not 0.5% (causes cell death)"
    └─ GPT-4o generates plan with correction already applied

No fine-tuning. No retraining. Expert knowledge compounds automatically.

Sample Hypotheses to Test

A paper-based electrochemical biosensor functionalized with anti-CRP antibodies
will detect C-reactive protein in whole blood at concentrations below 0.5 mg/L
within 10 minutes, matching laboratory ELISA sensitivity.

Supplementing C57BL/6 mice with Lactobacillus rhamnosus GG for 4 weeks will
reduce intestinal permeability by at least 30% compared to controls, measured
by FITC-dextran assay.

Replacing sucrose with trehalose as a cryoprotectant will increase post-thaw
viability of HeLa cells by at least 15 percentage points compared to the
standard DMSO protocol.

Name		Name	Last commit message	Last commit date
backend		backend
frontend		frontend
labprocure-video		labprocure-video
.gitignore		.gitignore
Justfile		Justfile
LICENSE		LICENSE
Product Video		Product Video
README.md		README.md
package-lock.json		package-lock.json
supabase_setup.sql		supabase_setup.sql

Navigation Menu

Repository navigation

About

Repository files navigation

LabProcure

From Hypothesis to Runnable Experiment with protocol design and lab procurement via Voice and Chat

What It Does

Tech Stack

Project Structure

Setup

1. Prerequisites

2. Clone and install

3. Configure environment

4. Set up Supabase (feedback loop only)

5. Run

API Endpoints

`POST /literature-qc`

`POST /generate-plan/stream`

`POST /chat`

`POST /voice-session`

`POST /feedback`

`POST /compare-material`

`POST /email-quote`

Voice Assistant — How It Works

Feedback Loop — How It Works

Sample Hypotheses to Test

License

About

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

LabProcure

From Hypothesis to Runnable Experiment with protocol design and lab procurement via Voice and Chat

What It Does

Tech Stack

Project Structure

Setup

1. Prerequisites

2. Clone and install

3. Configure environment

4. Set up Supabase (feedback loop only)

5. Run

API Endpoints

POST /literature-qc

POST /generate-plan/stream

POST /chat

POST /voice-session

POST /feedback

POST /compare-material

POST /email-quote

Voice Assistant — How It Works

Feedback Loop — How It Works

Sample Hypotheses to Test

License

About

Topics

Resources

Stars

Watchers

Forks

Contributors

Languages

`POST /literature-qc`

`POST /generate-plan/stream`

`POST /chat`

`POST /voice-session`

`POST /feedback`

`POST /compare-material`

`POST /email-quote`