Whitespace monitors AI research papers and uses an LLM pipeline to surface novel, feasible startup ideas hiding in the gaps between papers. Every run produces a ranked set of ideas with novelty and feasibility scores, detailed breakdowns, and a "product sketch" you can generate on demand.
Papers are fetched from arXiv, which is where leading AI research organisations — Google DeepMind, Anthropic, OpenAI, Meta AI, Mistral, and others — publish the majority of their work. Whitespace searches arXiv using configurable organisation names as keywords (e.g. all:DeepMind OR all:Anthropic) combined with subject category filters (e.g. cs.AI, cs.LG). This means any paper on arXiv that mentions a configured organisation in its title, abstract, or author affiliations is eligible for ingestion.
To pull in papers from a specific lab, add its name to ARXIV_ORGS in your .env. To focus on a particular research area, adjust ARXIV_CATEGORIES. Common examples:
| Organisation | Add to ARXIV_ORGS |
|---|---|
| Google DeepMind | DeepMind |
| Anthropic | Anthropic |
| OpenAI | OpenAI |
| Meta AI | Meta AI |
| Mistral | Mistral |
| Microsoft Research | Microsoft Research |
| Stanford HAI | Stanford |
| Berkeley AI Research | Berkeley |
arXiv papers
│
▼
Fetch Search arXiv for papers from configured orgs and categories
│
▼
Analyse LLM extracts key claims, methods, open questions per paper
│
▼
Gap map LLM identifies cross-paper research gaps and opportunities
│
▼
Synthesise LLM generates N startup ideas from the gap map
│
▼
Score LLM rates each idea on novelty (0–1) and feasibility (0–1)
│
▼
Select Top N ideas are persisted and tagged with the run ID
│
▼
Connect Ideas sharing source papers are linked to each other
Each pipeline run is saved independently so the History page accumulates every batch — you never lose earlier ideas when you run again.
Fast re-synthesis path: When no new arXiv papers are found, the pipeline skips the expensive per-paper LLM calls and builds lightweight pseudo-analyses directly from abstracts. Only two LLM calls are made (gap map + synthesise) instead of 30+, so subsequent runs on the same day complete in seconds rather than minutes.
| Layer | Technology |
|---|---|
| Backend API | FastAPI + SQLAlchemy (async) |
| Database | SQLite (dev) / PostgreSQL (prod) |
| Migrations | Alembic |
| Worker | Python threads, APScheduler |
| LLM runners | Claude CLI, Codex CLI, Gemini CLI, Anthropic API, Gemini API, OpenRouter |
| Frontend | React 18 + TypeScript + Vite |
| Data fetching | TanStack React Query |
| State | Zustand |
whitespace/
├── start.sh # One-command startup (installs deps, runs migrations, starts both servers)
├── backend/
│ ├── app/
│ │ ├── api/routes/ # FastAPI route handlers
│ │ │ ├── ideas.py # Today's feed, history, idea detail, surprise
│ │ │ ├── saved.py # Save / unsave ideas
│ │ │ ├── build.py # Trigger and fetch product sketch builds
│ │ │ ├── export.py # Export ideas as Markdown / PDF
│ │ │ └── system.py # Health, pipeline status/trigger, runner config, data sources
│ │ ├── db/
│ │ │ ├── models/ # SQLAlchemy ORM models
│ │ │ └── migrations/ # Alembic migration versions
│ │ ├── runners/ # LLM runner adapters (see Runner section)
│ │ ├── pipeline/ # Analysis, gap mapping, chunking, scoring utilities
│ │ └── core/config.py # Pydantic settings (reads .env)
│ └── worker/
│ ├── orchestrator.py # Full pipeline orchestration logic
│ ├── stages/ # fetch, analyse, gap_map, synthesise, score, select, connect
│ ├── prompts/ # Markdown prompt templates for each LLM stage
│ └── build_generator.py # Product sketch generation for saved ideas
└── frontend/
└── src/
├── pages/ # FeedPage, HistoryPage, IdeaDetailPage, SavedPage, SettingsPage, BuildOutputPage
├── components/ # NavBar, IdeaCard, HeroCard, BadgeRow, ScoreBar, ConnectedIdeas
├── hooks/ # useIdeas, useSaved, useBuild (React Query hooks)
└── api/ # Typed API client
git clone https://github.com/dgtise25/whitespace.git
cd whitespace
bash start.shThat single command:
- Stops any existing servers on ports 18730 / 18731
- Creates
backend/.envwith SQLite defaults if it doesn't exist - Creates and activates a Python virtualenv, installs all dependencies
- Installs frontend npm packages
- Runs Alembic migrations
- Starts the FastAPI backend on http://localhost:18730
- Starts the Vite frontend on http://localhost:18731
Open http://localhost:18731 in your browser, then click Refresh Ideas to run the pipeline.
./start.sh --dockerBuilds and starts Postgres, the FastAPI backend, and the Vite frontend via docker compose:
| Service | Container port | Host port |
|---|---|---|
| Postgres | 5432 | 18733 |
| Backend API | 8000 | 18730 |
| Frontend | 5173 | 18731 |
Config is read from the repo-root .env (see Configuration). Source lives under backend/ and frontend/ and is bind-mounted into the containers, so edits hot-reload without a rebuild.
The scheduled worker (daily pipeline run) is opt-in — it's defined behind a compose profile so it doesn't start by default:
docker compose -f docker/docker-compose.dev-minimal.yml --profile worker upTo run the compose stack directly instead of via start.sh:
docker compose -f docker/docker-compose.dev-minimal.yml up --buildCopy backend/.env.example to backend/.env and edit as needed:
# Database — SQLite for local dev, switch to postgres:// for productionDATABASE_URL=sqlite+aiosqlite:///./whitespace.db# LLM runner — configure at least one (see Runner section below)# ANTHROPIC_API_KEY=sk-ant-...# GEMINI_API_KEY=AIza...OPENROUTER_API_KEY=sk-or-...# OPENROUTER_ANALYSIS_MODEL=anthropic/claude-opus-5 # optional — pins an exact slug# Leave unset and Whitespace auto-resolves the newest Claude Opus model OpenRouter offers# Pipeline mode: "full" uses a real LLM; "stub" inserts fixture data (fast, no API calls)PIPELINE_MODE=full# Scheduled daily run time (24-hour clock, UTC)WORKER_SCHEDULE_HOUR=2WORKER_SCHEDULE_MINUTE=0# arXiv organisations to source papers fromARXIV_ORGS=DeepMind,Anthropic,OpenAI# arXiv subject categories to includeARXIV_CATEGORIES=cs.AI,cs.LG,cs.CL,cs.MA# Number of ideas to generate per pipeline runIDEAS_PER_RUN=8| Code | Subject |
|---|---|
cs.AI | Artificial Intelligence |
cs.LG | Machine Learning |
cs.CL | Computation and Language (NLP) |
cs.MA | Multi-Agent Systems |
cs.SE | Software Engineering |
cs.HC | Human-Computer Interaction |
eess.SP | Signal Processing |
Whitespace picks the first available runner in this priority order:
| Priority | Runner | How to enable |
|---|---|---|
| 1 | Claude CLI | Install the Claude Code CLI — no API key required |
| 2 | Codex CLI | Install the OpenAI Codex CLI |
| 3 | Gemini CLI | Install the Gemini CLI |
| 4 | Gemini API | Set GEMINI_API_KEY |
| 5 | Anthropic API | Set ANTHROPIC_API_KEY |
| 6 | OpenRouter | Set OPENROUTER_API_KEY |
You can override the active runner at runtime from the Settings page in the UI, or via the API:
# See which runners are available and which is active
curl http://localhost:18730/api/system/runners
# Pin to a specific runner
curl -X PUT http://localhost:18730/api/system/runner \
-H "Content-Type: application/json" \
-d '{"name": "anthropic"}'All endpoints are prefixed with /api. Interactive docs at http://localhost:18730/docs.
| Method | Path | Description |
|---|---|---|
GET | /api/ideas/today | Today's featured ideas (falls back to most recent run if none today) |
GET | /api/ideas/history | All ideas grouped by pipeline run, newest first |
GET | /api/ideas/surprise | Random featured idea |
GET | /api/ideas/{id} | Full idea detail including connected ideas |
Example — fetch today's feed:
curl http://localhost:18730/api/ideas/today{
"date": "2026-04-25",
"papers_ingested": 12,
"ideas": [
{
"id": "3fa85f64-...",
"title": "Federated Gap Detector",
"description": "A privacy-preserving system that surfaces research blind spots across siloed lab corpora without exposing raw data.",
"badge": "Novel",
"novelty_score": 0.91,
"feasibility_score": 0.74,
"is_featured": true,
"featured_date": "2026-04-25"
}
]
}Example — fetch idea detail:
curl http://localhost:18730/api/ideas/3fa85f64-...{
"id": "3fa85f64-...",
"title": "Federated Gap Detector",
"description": "...",
"why_novel": "No existing tool combines federated learning with cross-corpus gap analysis.",
"who_builds": "ML infrastructure teams at research-heavy organisations.",
"who_buys": "AI labs, pharma companies, government research bodies.",
"novelty_score": 0.91,
"feasibility_score": 0.74,
"badge": "Novel",
"paper_ids": ["2404.12345", "2404.67890"],
"connections": [
{
"id": "abc-...",
"title": "Cross-Silo Knowledge Distillation",
"badge": "Feasible",
"shared_paper_count": 2
}
]
}| Method | Path | Description |
|---|---|---|
GET | /api/saved/ | List all saved ideas |
POST | /api/saved/ | Save an idea {"idea_id": "..."} |
DELETE | /api/saved/{idea_id} | Remove a saved idea |
Example:
# Save an idea
curl -X POST http://localhost:18730/api/saved/ \
-H "Content-Type: application/json" \
-d '{"idea_id": "3fa85f64-..."}'# List saved ideas
curl http://localhost:18730/api/saved/| Method | Path | Description |
|---|---|---|
GET | /api/build/{idea_id} | Fetch existing product sketch |
POST | /api/build/{idea_id} | Trigger product sketch generation (async, returns 202) |
Example:
# Trigger a build
curl -X POST http://localhost:18730/api/build/3fa85f64-...
# Poll until ready (status changes from "generating" to "ready")
curl http://localhost:18730/api/build/3fa85f64-...{
"idea_id": "3fa85f64-...",
"status": "ready",
"product_sketch": {
"tagline": "Find the gaps your competitors can't see.",
"target_user": "Research leads at AI labs",
"core_loop": "Ingest → Analyse → Surface → Act",
"risks": [
{ "title": "Data access", "description": "Labs may not share paper corpora." }
],
"monetisation": [
{ "name": "SaaS subscription", "fit": "high", "description": "Per-seat pricing for research teams." }
]
}
}| Method | Path | Description |
|---|---|---|
GET | /api/export/{idea_id}/markdown | Download idea + build as a .md file |
GET | /api/export/{idea_id}/pdf | Download idea + build as a .pdf file |
curl http://localhost:18730/api/export/3fa85f64-.../markdown -o idea.md
curl http://localhost:18730/api/export/3fa85f64-.../pdf -o idea.pdf| Method | Path | Description |
|---|---|---|
GET | /api/system/health | Health check — API and database status |
GET | /api/system/pipeline/status | Whether pipeline is running + last completed run |
POST | /api/system/pipeline/run | Trigger a pipeline run manually |
GET | /api/system/runners | List available LLM runners and active runner |
PUT | /api/system/runner | Set preferred runner |
GET | /api/system/config | Current data source configuration |
PUT | /api/system/data-sources | Update active orgs and categories |
Example — trigger the pipeline:
curl -X POST http://localhost:18730/api/system/pipeline/run{ "status": "started", "message": "Pipeline started in background." }Example — update data sources:
curl -X PUT http://localhost:18730/api/system/data-sources \
-H "Content-Type: application/json" \
-d '{"orgs": ["DeepMind", "Anthropic"], "categories": ["cs.AI", "cs.CL"]}'| Page | Route | Description |
|---|---|---|
| Ideas | / | Today's featured ideas — hero card for the top idea, grid below |
| History | /history | Every pipeline run accumulated over time, filterable by badge |
| Idea detail | /ideas/:id | Full breakdown: why novel, who builds, who buys, connected ideas |
| Build | /ideas/:id/build | AI-generated product sketch for a specific idea |
| Saved | /saved | Ideas you've bookmarked |
| Settings | /settings | Configure LLM runner, arXiv orgs, and categories |
The NavBar polls the pipeline status every 8 seconds. When a run completes, the Ideas and History pages refresh automatically — no manual reload needed.
| Table | Purpose |
|---|---|
papers | Raw arXiv papers (title, abstract, authors, categories, URL) |
chunks | Text chunks derived from paper abstracts |
ingestion_runs | One row per pipeline run — tracks papers fetched, ideas generated, errors |
ideas | Generated ideas with scores, badges, and run_id linking to the ingestion run |
connected_ideas | Pairs of ideas that share source papers, ranked by shared paper count |
saved_ideas | User bookmarks linking to ideas |
build_outputs | AI-generated product sketches for saved ideas |
Each idea receives one badge based on its scores:
| Badge | Meaning |
|---|---|
| Novel | High novelty (≥ 0.7), lower feasibility — forward-looking research opportunity |
| Feasible | High feasibility (≥ 0.7), lower novelty — buildable with current technology |
| Emerging | Both scores moderate — interesting but early |
| Speculative | Both scores lower — long-horizon, high-risk |
Switch to PostgreSQL by updating DATABASE_URL:
DATABASE_URL=postgresql+asyncpg://user:password@localhost:5432/whitespaceRun migrations with the synchronous driver:
DATABASE_URL=postgresql+psycopg://user:password@localhost:5432/whitespace \
alembic upgrade headStart the backend with a production ASGI server:
uvicorn app.main:app --host 0.0.0.0 --port 18730 --workers 2Build the frontend for production:
cd frontend && npm run build
# Serve the dist/ folder with nginx or any static host# Backend testscd backend && pytest
# Frontend testscd frontend && npm test# Type-check frontendcd frontend && npm run build
# Lint backendcd backend && ruff check .To use the stub pipeline (no LLM calls, instant fixture data):
PIPELINE_MODE=stub