TARBoard — Targeted Award Radar for Biomedical Oncology Research

A single-file web dashboard that matches a researcher's CV + research topics against U.S. funding opportunities in AI / computational oncology / blood cancer, and ranks them by fit. Built for a senior postdoc planning the transition to independence.

Two-stage engine: deterministic rules decide eligibility (citizenship / PR, post-PhD window, MD-only, nationality — auditable and never hallucinated), and Claude provides the judgment on top — it reads your CV into a structured profile, re-ranks the matches with a reason for each, reads federal eligibility prose, and drafts a CV-grounded pitch angle for any grant. Every Claude output is grounded strictly on the retrieved data (it never invents a grant, deadline, or amount), and the whole app degrades gracefully to lexical matching when Claude isn't connected. Built with Claude Code.

Open it

Just open index.html in a browser (double-click, or open index.html). No build step or server needed to use it.

• If your browser restricts the live federal search over file://, serve locally instead:

  python3 -m http.server 8912 --directory .
  # then visit http://localhost:8912/index.html

What it does

• Upload a CV (PDF or text) or paste it — keywords are extracted client-side. Click ★ Load example — Jordan Bennett to try it with an example profile.
• Add research topics (blood-cancer subtypes, AI methods) as chips.
• Search returns opportunities ranked by a match score, each showing award size, deadline (with day-countdown), eligibility, career stage, U.S.-citizen flags, and why it matched.
• Save opportunities to a shortlist (persisted in your browser) and export to CSV/JSON.
• Light/dark themes.

Connecting Claude

The dashboard is a static file, so it can't hold an API key. Two ways to give it one:

1. Cloudflare Worker proxy (recommended). Deploy the ~90-line proxy in worker/ (holds the key as a secret; model allowlist + CORS + optional per-IP cap). See worker/README.md — it's a 2-minute wrangler deploy. Then click the ⚙ button in the app and paste the Worker URL.
2. Paste-your-own-key dev mode. For local rehearsal only: ⚙ → paste an Anthropic key (stays in your browser, calls the API directly). Never ship a page with a key baked in — build.py refuses to build if one is present.
3. Claude Code bridge (no key, uses your Claude subscription). For a local demo, run python3 worker/claude_code_bridge.py and point ⚙ at http://localhost:8788; it runs the app's Claude calls through the claude -p CLI on your existing login. See worker/README.md. Local-only and slower than the API, but zero setup — good for developing/demoing on a laptop.

Without either, the app still runs fully on the deterministic/lexical path. Models are pinned in one CONFIG object (claude-opus-4-8 workhorse, claude-haiku-4-5 fast path); ~$0.15–0.25 per session.

Data sources

Source	Count	How
Grants.gov (NIH, NSF, DOD/CDMRP, ARPA-H)	live	Queried in real time from the official Search2 API
Curated foundations/societies/industry	187	data/curated_opportunities.json — see FOUNDATIONS.md
JHU list (refreshed 2026)	81	Seeded from the JHU data/JHU_*.xlsx files, then every program's deadline/amount/eligibility was web-verified to 2026-2027 (data/jhu_updated.json); discontinued programs dropped

Eligibility & sorting: the top bar takes PhD year and citizenship into account. Results default to soonest-deadline order (upcoming first; closed ones sink to the bottom), match your career stage, and — for profiles without U.S. citizenship/permanent residence — hide awards that require it (counted separately; reveal with "Show citizenship-locked"). Awards past a hard post-PhD window (e.g. K99's 4-year limit) are flagged.

FOUNDATIONS.md is the human-readable catalog of the curated non-federal funders (computational / blood cancer / rare disease), grouped by theme with senior-postdoc guidance. It and the dashboard's "Curated foundations" source are the same dataset.

Rebuilding after data changes

The dataset JSON is inlined into index.html at build time from index.template.html:

python3 build.py     # regenerates index.html from the template + data/*.json + user_profile CV

Edit index.template.html for app/UI changes and data/*.json for data changes, then re-run build.py.

Deploy the page

The live app is a Cloudflare Pages project (tarboard → https://tarboard.pages.dev), deployed as a direct upload of the single built file — not connected to the Git repo. After build.py, copy index.html into an otherwise-empty directory and deploy that directory:

python3 build.py
mkdir -p dist && cp index.html dist/          # dist/ holds ONLY index.html
npx wrangler pages deploy dist --project-name tarboard --branch main

⚠️ Never deploy the repo root. Point wrangler pages deploy at a directory containing only the built index.html. The repo root holds user_profile/ (real CV) and other draft material that is fine on disk but must never be published — deploying . would upload all of it. main is the production branch, so --branch main publishes to the tarboard.pages.dev apex; any other branch name creates a throwaway preview URL instead.

Files

• index.html — the built, self-contained app (open this)
• index.template.html — source template with data placeholders
• build.py — inlines data into the template (and asserts no API key is baked in)
• worker/ — Cloudflare Worker Claude proxy + deploy guide
• IMPLEMENTATION_PLAN.md — the Claude-integration build plan / spec
• FOUNDATIONS.md — curated funder catalog (127 non-federal programs + federal appendix)
• data/curated_opportunities.json, data/jhu_updated.json — normalized data
• data/federal_seed.json — real Grants.gov snapshot used as an offline fallback
• user_profile/ — example CV

⚠️ Award amounts, deadlines, and eligibility change frequently. Always confirm on the funder's own page before applying.