Vermont Research Reviewer

A dual-mode multi-agent review tool built on the Anthropic Claude API. Submit a research proposal or project outline and receive structured feedback from expert reviewers simultaneously, synthesized into an actionable summary.

Two audiences, one tool:

• Student mode — scaffolded feedback for ORCA student interns at UVM
• Advisor mode — direct critical analysis for faculty, program managers, and grant reviewers

---

What It Does

Each analysis runs one parallel Claude API call per selected reviewer role, each grounded in an agency or program reference document. A final synthesis call consolidates the feedback into a structured summary tuned to the selected mode.

Reviewer Roles

Reviewer	Focus
Research Methodologist	Study design, IRB requirements, scope feasibility, data collection validity
VT Dept of Environmental Conservation (DEC)	Act 250, stormwater/wetland/wastewater permits, water quality, groundwater, PFAS
VT Agency of Natural Resources (ANR)	Wildlife habitat, biodiversity, rare species, forest ecology, environmental justice
VT Dept of Health (VDH)	Human subjects, HIPAA, health data access, health equity
VT Agency of Agriculture, Food & Markets (AAFM)	Required Agricultural Practices, farm engagement, water/ag nexus, food safety
NSF Program Officer	Intellectual Merit, Broader Impacts, data management, EPSCoR, budget compliance
NIH Program Officer	Review criteria, human subjects, rigor/reproducibility, Vermont health data risks
Sloan Foundation Program Officer	Program fit, scientific rigor, public good value, measurable outcomes, LOI-first process, budget-scope alignment
Wellcome Program Officer	Bold/creative research criteria, eligibility gates, data sharing mandate, LMIC access, budget justification
UVM IRB Specialist	IRB classification, CITI training, consent, HIPAA/FERPA, data security
UVM Sponsored Research Officer	OSP sign-off, F&A rates, Uniform Guidance, export control, subawards

Each role has two system prompt variants — same context document, different tone and framing depending on mode.

Reviewers are individually selectable via checkboxes — none are selected by default. Select at least one to run an analysis.

Synthesis Output

Student mode produces:

1. Critical Issues to Resolve Before Starting
2. Compliance Steps Required
3. Key Risks to Manage
4. Domain Knowledge Gaps
5. Questions to Bring to Your Faculty Advisor

Advisor mode produces:

1. Where Reviewers Agree
2. Key Conflicts and Tensions
3. Regulatory and Compliance Exposure
4. Strategic Gaps and Blind Spots
5. Recommended Next Steps

---

Project Structure

agentic-team/
├── app.py                        # Streamlit web UI (main entry point)
├── requirements.txt
├── .env                          # Your API key (never commit this)
├── .env.example                  # Safe committed placeholder
├── CHANGELOG.md                  # Design evolution log
├── persona_context/
│   ├── research_methodologist_context.md
│   ├── vt_dec_context.md
│   ├── vt_anr_context.md
│   ├── vt_health_context.md
│   ├── vt_agriculture_context.md
│   ├── nsf_program_officer_context.md
│   ├── nih_program_officer_context.md
│   ├── sloan_program_officer_context.md
│   ├── wellcome_program_officer_context.md
│   ├── uvm_irb_context.md
│   └── uvm_sponsored_research_context.md

---

Setup

1. Clone and install dependencies

git clone <repo-url>
cd agentic-team
pip install -r requirements.txt

2. Get your Anthropic API key

• Go to console.anthropic.com
• Navigate to Settings → API Keys → Create Key
• Add at least $5 in credits under Settings → Billing

3. Configure your key

Copy .env.example to .env and add your key:

cp .env.example .env

Edit .env:

ANTHROPIC_API_KEY=sk-ant-...

Never commit .env — it is listed in .gitignore.

4. Run the web UI

streamlit run app.py

Opens at http://localhost:8501.

5. Stop the server

Press Ctrl+C in the terminal where Streamlit is running.

If the process is stuck or running in the background:

# Find the process
lsof -i :8501

# Kill it by PID
kill <PID>

---

Mode Selection

The sidebar toggle switches between two modes:

Mode	Audience	Reviewer tone	Synthesis structure
Student	ORCA interns	Explains concepts, flags unknowns	Blockers → compliance → risks → gaps → advisor questions
Advisor	Faculty, program managers, grant reviewers	Direct, assumes domain fluency	Agreement → conflicts → compliance exposure → blind spots → next steps

Model defaults change automatically with mode but can be overridden.

---

Model Selection

Option	Model	Best for
Haiku (fast, low cost)	claude-haiku-4-5-20251001	Development, student mode
Sonnet (balanced)	claude-sonnet-4-6	Default production use
Opus (most capable)	claude-opus-4-7	High-stakes proposals

Student mode defaults: Haiku for reviewers, Sonnet for synthesis. Advisor mode defaults: Sonnet for reviewers, Sonnet for synthesis.

Estimated cost per analysis: ~$0.02–0.10 depending on mode, model, and input length.

---

Adding or Updating Context Documents

Each reviewer role loads a markdown reference file from persona_context/. To improve reviewer quality:

1. Edit the relevant .md file in persona_context/
2. Add specific local knowledge: permit numbers, agency contacts, program names, known constraints
3. Keep files under ~8,000 tokens for best cost/performance ratio

To add a new reviewer role:

1. Create a new context file in persona_context/
2. Add an entry to the ROLES dict in app.py — include both student and advisor prompt variants
3. The UI will automatically render a checkbox for the new reviewer

---

Web Fetching

The UI accepts optional reference URLs (one per line). Each URL is fetched at run time, stripped of navigation/scripts, and appended to every reviewer's user message. Capped at 8,000 characters per URL. Sites that block automated access return a graceful error rather than crashing.

---

Security

• API key is stored in .env and loaded with override=True to prevent Windows system environment variables from shadowing it
• .env is gitignored — never commit it
• No user data is stored or logged by this application
• All API calls go directly to Anthropic — no third-party services