Production-ready autonomous grant discovery and proposal generation system with MCP scrapers, A2A orchestration, multi-model AI agents, and real-time visualization.
This is a complete grant application automation platform that:
- Discovers relevant grants from multiple sources (EU Horizon, NSF, private foundations)
- Matches grants to your company profile using AI-powered relevance scoring
- Generates proposals in two modes:
- ⚡ Fast Track (30 seconds) - Instant proposals via Gemini 2.5 Pro
- 🔬 Research Track (5-15 minutes) - Deep autonomous research via AI Agent Laboratory
The system follows a linear 4-phase pipeline, designed for autonomy and precision:
graph LR
%% PHASE 1: DISCOVERY
subgraph PHASE_1 [PHASE 1: DISCOVERY & MATCHING]
direction TB
Sources[Sources: EU/NSF/Foundations] --> Firecrawl[Firecrawl Engine]
Firecrawl --> Analyzer[AI Feature/Relevance Extractor]
Analyzer -->|Score > 50| DB[(Supabase DB)]
Analyzer -->|Score < 50| Discard[Discard]
end
%% PHASE 2: INTERFACE
subgraph PHASE_2 [PHASE 2: PLATFORM]
DB <==>|Real-time| Dashboard[SvelteKit Dashboard]
Dashboard -->|User Clicks Apply| API[API Gateway]
end
%% PHASE 3: ORCHESTRATION
subgraph PHASE_3 [PHASE 3: ORCHESTRATION]
API --> Strategy{Strategy Router}
Strategy -->|Fast Track| Gemini[Gemini 2.5 Pro]
Strategy -->|Research Track| LabRunner[Lab Orchestrator]
end
%% PHASE 4: AGENT LAB
subgraph PHASE_4 [PHASE 4: AGENT LABORATORY]
direction TB
LabRunner --> PhD[PhD Student - Claude Opus 4.5]
PhD --> Postdoc1[Postdoc Plan - Gemini 3 Pro]
Postdoc1 --> Eng[ML and SW Engineers - GPT-5 Codex]
Eng --> Postdoc2[Results Analysis - Gemini 3 Pro]
Postdoc2 --> Prof[Professor Writing - Claude Sonnet 4.5]
Prof --> Review[Review Board - Gemini 3 Pro]
Review --> Final[Research-Grade Proposal]
end
- Sources: Continuously monitors EU Horizon, grants.gov, and private foundations.
- Firecrawl: Scrapes raw unstructured text from grant pages into Markdown.
- AI Filter: Uses Gemini Pro to extract standard metadata and calculate a Relevance Score (0-100) specific to your startup's profile. Only high-value matches enter the database.
Once a user selects "Apply", the system routes the request based on depth needs:
- Fast Track: Uses a single-shot prompt with Gemini 2.5 Pro for instant (~30s) drafts.
- Research Track: Activates the Agent Laboratory (below) for deep investigation.
The Research Track uses a team of 6 specialized AI agents, each running on the optimal model for their task:
╔══════════════════════════════════════════════════════════════════════════════════════╗
║ 🔬 AI RESEARCH LABORATORY WORKFLOW ║
║ Maximum Quality Multi-Model Orchestration ║
╠══════════════════════════════════════════════════════════════════════════════════════╣
║ ║
║ PHASE 1: LITERATURE REVIEW ║
║ ┌─────────────────────────────────────────────────────────────────────────────┐ ║
║ │ 📚 PhD Student Agent │ ║
║ │ ───────────────────────────────────────────────────────────────────────── │ ║
║ │ Model: Claude Opus 4.5 (Reasoning) Cost: ~$2.50/run │ ║
║ │ Why: Best reasoning (69.77 score), excellent research synthesis │ ║
║ │ │ ║
║ │ Tasks: │ ║
║ │ • Search academic papers related to grant topic │ ║
║ │ • Analyze 10-20 relevant publications │ ║
║ │ • Extract key methodologies and findings │ ║
║ │ • Identify research gaps and opportunities │ ║
║ │ │ ║
║ │ Output: Literature review summary, cited sources, key insights │ ║
║ └─────────────────────────────────────────────────────────────────────────────┘ ║
║ │ ║
║ ▼ ║
║ PHASE 2: PLAN FORMULATION ║
║ ┌─────────────────────────────────────────────────────────────────────────────┐ ║
║ │ 📋 Postdoc Agent │ ║
║ │ ───────────────────────────────────────────────────────────────────────── │ ║
║ │ Model: Gemini 3 Pro (high) Cost: FREE ✨ │ ║
║ │ Why: Top reasoning (72.85 score), 95.67% math, FREE via Gemini API │ ║
║ │ │ ║
║ │ Tasks: │ ║
║ │ • Synthesize literature findings │ ║
║ │ • Formulate research methodology │ ║
║ │ • Define project milestones and deliverables │ ║
║ │ • Create detailed research plan │ ║
║ │ │ ║
║ │ Output: Research plan, methodology, timeline, success metrics │ ║
║ └─────────────────────────────────────────────────────────────────────────────┘ ║
║ │ ║
║ ▼ ║
║ PHASE 3: DATA PREPARATION ║
║ ┌─────────────────────────────────────────────────────────────────────────────┐ ║
║ │ 📊 ML Engineer Agent + 🔧 SW Engineer Agent │ ║
║ │ ───────────────────────────────────────────────────────────────────────── │ ║
║ │ ML Model: GPT-5 Codex (high) Cost: ~$1.00/run │ ║
║ │ SW Model: Claude Opus 4.5 Cost: ~$2.00/run │ ║
║ │ Why: GPT-5 Codex has 98.67% math, Claude has 87.1% LiveCodeBench │ ║
║ │ │ ║
║ │ Tasks (ML Engineer): │ ║
║ │ • Identify required datasets │ ║
║ │ • Design data collection strategy │ ║
║ │ • Prepare data processing pipelines │ ║
║ │ │ ║
║ │ Tasks (SW Engineer): │ ║
║ │ • Set up development environment │ ║
║ │ • Implement data loaders and transformations │ ║
║ │ • Ensure code quality and documentation │ ║
║ │ │ ║
║ │ Output: Dataset specifications, code implementations, requirements.txt │ ║
║ └─────────────────────────────────────────────────────────────────────────────┘ ║
║ │ ║
║ ▼ ║
║ PHASE 4: RUNNING EXPERIMENTS ║
║ ┌─────────────────────────────────────────────────────────────────────────────┐ ║
║ │ 🧪 ML Engineer Agent (continued) │ ║
║ │ ───────────────────────────────────────────────────────────────────────── │ ║
║ │ Model: GPT-5 Codex (high) Cost: (included above) │ ║
║ │ │ ║
║ │ Tasks: │ ║
║ │ • Execute experimental protocols │ ║
║ │ • Run simulations and benchmarks │ ║
║ │ • Collect and validate results │ ║
║ │ • Generate visualizations │ ║
║ │ │ ║
║ │ Output: Experimental results, charts, performance metrics │ ║
║ └─────────────────────────────────────────────────────────────────────────────┘ ║
║ │ ║
║ ▼ ║
║ PHASE 5: RESULTS INTERPRETATION ║
║ ┌─────────────────────────────────────────────────────────────────────────────┐ ║
║ │ 📈 Postdoc Agent (continued) │ ║
║ │ ───────────────────────────────────────────────────────────────────────── │ ║
║ │ Model: Gemini 3 Pro (high) Cost: FREE ✨ │ ║
║ │ │ ║
║ │ Tasks: │ ║
║ │ • Analyze experimental outcomes │ ║
║ │ • Compare results to initial hypotheses │ ║
║ │ • Identify key findings and implications │ ║
║ │ • Assess alignment with grant objectives │ ║
║ │ │ ║
║ │ Output: Results analysis, key findings, comparison to objectives │ ║
║ └─────────────────────────────────────────────────────────────────────────────┘ ║
║ │ ║
║ ▼ ║
║ PHASE 6: REPORT WRITING ║
║ ┌─────────────────────────────────────────────────────────────────────────────┐ ║
║ │ 📝 Professor Agent │ ║
║ │ ───────────────────────────────────────────────────────────────────────── │ ║
║ │ Model: Claude Sonnet 4.5 (Reasoning) Cost: ~$1.50/run │ ║
║ │ Why: Excellent writing quality + reasoning balance │ ║
║ │ │ ║
║ │ Tasks: │ ║
║ │ • Structure the grant proposal │ ║
║ │ • Write executive summary │ ║
║ │ • Compose methodology section │ ║
║ │ • Draft budget justification │ ║
║ │ • Create impact statement │ ║
║ │ │ ║
║ │ Output: Complete grant proposal draft in LaTeX/Markdown │ ║
║ └─────────────────────────────────────────────────────────────────────────────┘ ║
║ │ ║
║ ▼ ║
║ PHASE 7: QUALITY REVIEW & REFINEMENT ║
║ ┌─────────────────────────────────────────────────────────────────────────────┐ ║
║ │ ✨ Reviewers Panel │ ║
║ │ ───────────────────────────────────────────────────────────────────────── │ ║
║ │ Model: Gemini 3 Pro (high) Cost: FREE ✨ │ ║
║ │ Why: Top GPQA (90.8%), rigorous logical reasoning │ ║
║ │ │ ║
║ │ Tasks: │ ║
║ │ • Score proposal on multiple criteria │ ║
║ │ • Identify weaknesses and gaps │ ║
║ │ • Suggest specific improvements │ ║
║ │ • Verify alignment with grant requirements │ ║
║ │ • Final quality assessment │ ║
║ │ │ ║
║ │ Output: Review scores, improvement suggestions, final approval │ ║
║ └─────────────────────────────────────────────────────────────────────────────┘ ║
║ │ ║
║ ▼ ║
║ ┌─────────────────────────────────────────────────────────────────────────────┐ ║
║ │ 📄 FINAL OUTPUT: RESEARCH-GRADE PROPOSAL │ ║
║ │ │ ║
║ │ • Complete grant proposal with citations │ ║
║ │ • Detailed methodology and timeline │ ║
║ │ • Budget breakdown and justification │ ║
║ │ • Impact statement and broader implications │ ║
║ │ • Quality score and confidence rating │ ║
║ └─────────────────────────────────────────────────────────────────────────────┘ ║
║ ║
╠══════════════════════════════════════════════════════════════════════════════════════╣
║ 💰 COST SUMMARY (per research run) ║
║ ─────────────────────────────────────────────────────────────────────────────────── ║
║ PhD Student (Claude Opus 4.5): $2.50 ║
║ Postdoc (Gemini 3 Pro): $0.00 ← FREE! ║
║ ML Engineer (GPT-5 Codex): $1.00 ║
║ SW Engineer (Claude Opus 4.5): $2.00 ║
║ Professor (Claude Sonnet 4.5): $1.50 ║
║ Reviewers (Gemini 3 Pro): $0.00 ← FREE! ║
║ ─────────────────────────────────────────────────────────────────────────────────── ║
║ TOTAL ESTIMATED COST: $7.00 - $15.00 per proposal ║
╚══════════════════════════════════════════════════════════════════════════════════════╝
- Node.js 18+
- Python 3.10+ (for Research Track)
- Supabase account (free tier works)
- API Keys:
- Gemini API (free, required)
- OpenRouter API (for Maximum Quality mode, ~$20 credit recommended)
cd backend
npm install
# Copy environment template
cp .env.example .env
# Edit .env with your API keys:
# - SUPABASE_URL
# - SUPABASE_ANON_KEY
# - GEMINI_API_KEY (free from Google AI Studio)
# - OPENROUTER_API_KEY (for Maximum Quality mode)
npm run devNote: For production deployment, you must set
PUBLIC_API_URLin your frontend environment (e.g., Vercel) to point to the live public URL of the deployed backend service (e.g.,https://your-backend.up.railway.app/api).
Database Setup: Run schema.sql in your Supabase SQL editor.
cd frontend
npm install
npm run devWindows Users: Requires Microsoft Visual C++ 14.0+ Build Tools
# Install from: https://visualstudio.microsoft.com/visual-cpp-build-tools/Then for all platforms:
cd backend/ai-researcher
pip install -e .
playwright installai-grant-crawler-a2a-pro/
│
├── backend/ # Node.js/Express API Server
│ ├── src/
│ │ ├── routes/ # API endpoints
│ │ │ ├── grants.js # Grant CRUD operations
│ │ │ └── proposals.js # Proposal generation endpoints
│ │ ├── services/
│ │ │ ├── grantMatcher.js # AI relevance scoring
│ │ │ └── aiResearcher.js # Research Track orchestration
│ │ ├── config/
│ │ │ └── gemini.js # Gemini 3 Pro configuration
│ │ └── utils/
│ │ └── sseHelper.js # Server-Sent Events streaming
│ │
│ └── ai-researcher/ # Python AI Research Laboratory
│ ├── agents.py # Agent class definitions
│ ├── agent_models.py # Per-agent model configuration ✨ NEW
│ ├── inference.py # Multi-provider API (OpenRouter, Gemini, OpenAI)
│ ├── ai_lab_repo.py # Main workflow orchestrator
│ └── experiment_configs/ # YAML experiment configurations
│
├── frontend/ # SvelteKit Application
│ ├── src/
│ │ ├── routes/
│ │ │ ├── +page.svelte # Grant dashboard
│ │ │ └── thinktank/
│ │ │ └── [grantId]/ # Research visualization page
│ │ └── lib/
│ │ └── components/ # Reusable UI components
│ │
│ └── static/ # Static assets
│
└── docs/ # Documentation & diagrams
| Key | Provider | Cost | Used For |
|---|---|---|---|
GEMINI_API_KEY |
Google AI Studio | FREE | Postdoc, Reviewers, Fast Track |
| Key | Provider | Cost | Used For |
|---|---|---|---|
OPENROUTER_API_KEY |
OpenRouter | ~$20+/run | Claude Opus 4.5, GPT-5 Codex, Claude Sonnet 4.5 |
The model selections are based on December 2024-2025 benchmark data from:
- Artificial Analysis Leaderboard
- AIME 2025, GPQA Diamond, LiveCodeBench, HumanEval
This system implements the Agent Laboratory framework as detailed in the paper Agent Laboratory: Using LLM Agents as Research Assistants (Schmidgall et al., 2025).
The architecture faithfully reproduces the paper's three-phase autonomous research workflow:
- Literature Review Phase: Implements the paper's "PhD Student" agent role to query arXiv/Semantic Scholar, performing independent literature synthesis rather than simple summarization.
- Experimentation Phase: Utilizes the MLE-Solver methodology described in the paper, where an ML Engineer agent iteratively refines code based on error traces and performance metrics (achieving state-of-the-art performance on benchmarks).
- Report Writing Phase: Adopts the "Professor" agent persona to synthesize findings into a coherent narrative, separating ideation (human/agent collaboration) from the labor of drafting.
Research Impact: The original study demonstrated that this multi-agent approach reduces research costs by 84% compared to traditional methods while maintaining or exceeding human-level quality in comparable tasks.
MIT License - feel free to use this for your grant applications!
- Agent Laboratory - Original research agent framework
- Anthropic, Google, OpenAI, DeepSeek - For the amazing AI models
- You - For building the future of automated grant writing!