Open-source YC batch analyzer. One command produces an evidence-grounded VC-style research report and dashboard for the most recent Y Combinator batch.
The dashboard auto-refuses to ship if any cited URL is dead. Full-page screenshot at docs/screenshots/dashboard-w26.jpg.
yc-ai-pulse answers: what does the most recent YC batch tell us about the state of AI?
For every company in the latest batch, it classifies:
- Industry — primary + secondary
- AI capability — code-gen, agents, RAG, voice, vision, robotics, ...
- Tech stack — model providers, frameworks, infra
- OSS posture — fully-open, weights-only, api-only, closed
Then it produces three artifacts, all grounded in a single CSV — no hallucinated numbers, every quote traceable, every link verified before publish:
- Interactive HTML dashboard with drill-down to source rows
- VC-style PowerPoint deck (a16z aesthetic) —
.pptx - Narrative memo —
.docx
A Chrome extension wraps two flows: (a) analyze the whole batch, (b) deep-dive a single company.
| Phase | Surface | Status |
|---|---|---|
| 0 | Repo bootstrap, secrets hygiene, CI | ✅ shipped |
| 1 | CLI + dashboard with anti-hallucination Layer 1 | ✅ v0.1.0 |
| 2 | Depth=1 crawler + ECharts dashboard + .pptx deck + .docx memo |
✅ shipped |
| 3 | Chrome extension | ⬜ planned |
See CHANGELOG.md for what 0.1.0 includes, BACKLOG.md for the working backlog, and docs/decisions/ for architecture decisions.
pipx install yc-ai-pulse # or: uv tool install yc-ai-pulse
# Coverage probe only (no LLM cost) — fetches the latest batch and shows
# what's analyzable. Headline: % of YC batch covered, with the dropped
# register naming every excluded company.
ycai run-coverage --batch winter-2026 --yc-official-count 196
# Full enrichment via your Claude Max subscription (~6 min on W26, ~free).
# 95% high-confidence rate. Renders the dashboard with capability heatmap,
# tech-stack and OSS-posture breakdowns. Refuses to write the dashboard
# if any cited URL is dead.
ycai run-coverage --batch winter-2026 --yc-official-count 196 --enrich
# Pay-per-token instead of subscription:
export ANTHROPIC_API_KEY=sk-ant-...
ycai run-coverage --batch winter-2026 --yc-official-count 196 --enrich
# Resume an interrupted run (quota wall, crash, network blip):
ycai resume runs/2026-05-01-XXXXXX
# Re-render the dashboard from existing artifacts at zero LLM cost
# (useful when the dashboard layout changes):
ycai dashboard runs/2026-05-01-XXXXXXA real run on YC W26 is checked in as a working example: see examples/output/dashboard-w26-pr4-2026-05-01.html. The full quality writeup is in docs/QUALITY_REPORT_W26.md.
This tool is built around the rule that numbers come from pandas, never from the LLM. Concretely:
- Schema-enforced LLM output — pydantic models reject any row without ≥1 source URL.
- Two-pass cross-check on uncertain classifications; disagreements drop out of charts.
- Numerical drift check — any number in generated prose that's not in the underlying CSV fails the build.
- Forbidden-phrase scanner — "studies show", "experts say", and similar lazy hedges fail the build.
- Quote-span verification — every quote re-fetched from source at report time, byte-compared.
- Link-verify hard gate — no artifact ships if any cited link returns 4xx/5xx.
Each chart in the dashboard exposes the source rows that produced it. If the data isn't there, the chart doesn't render.
yc-ai-pulse is local-only. No telemetry, no remote config, no data leaves your machine except for the LLM calls (to Anthropic) and the public-page fetches you'd make in a browser.
The pipeline collects only data already public on yc.com. A sanitizer.py strips emails, phone numbers, and addresses defensively before any data hits disk or the LLM.
MIT — see LICENSE.