v0.2.0 — Phase 2 (crawler + ECharts + deck + memo)

Phase 2 release. Same end-to-end pipeline as 0.1.0 but with materially richer output:

What's new

• Depth=1 polite website crawl (PR #12). Robots-aware, max 5 pages × 30 KB × 4-second timeout per company. OSS posture unknown rate dropped from 55% to 21% on the same W26 cohort. Tech-stack identified mentions: 14 → 41.
• Apache ECharts replaces the dashboard's static CSS bars (PR #13). Real interactive heatmap with hover tooltips, pies with click-to-isolate, axis-pointer bar charts. Pure JSON options (no client-side eval); SHA-384 SRI hash; <noscript> fallback to source-row tables.
• VC-style .pptx deck (PR #14). 16 slides, a16z-feel cream/orange palette. matplotlib chart PNGs anchored to a shared analytics.py module. Anti-hallucination Layer 2 runs before write: forbidden-phrase scan + numerical-drift check.
• Narrative .docx memo (PR #15). 9 sections per USER.md document-format discipline. Same chart math as the deck. Dual-output ycai report <run-dir> produces both deck.pptx and report.docx.

W26 dogfood

• 113/124 high-confidence (91%) on the LLM enrichment, 0 hallucinated source URLs
• 65 of 113 high-confidence companies build agents (58% of the cohort) — defensible with row-level evidence in the heatmap
• 21% OSS-posture unknown rate (down from 55% in 0.1.0)
• Layer 2 audit ran clean on the real run

Quickstart

pipx install yc-ai-pulse
ycai run-coverage --batch winter-2026 --yc-official-count 196 --enrich
ycai dashboard runs/2026-*/        # interactive HTML
ycai report runs/2026-*/           # produces deck.pptx + report.docx

Sample artifacts (in repo)

• examples/output/dashboard-w26-pr12-2026-05-01.html
• examples/output/deck-w26-pr14-2026-05-01.pptx
• examples/output/report-w26-pr15-2026-05-01.docx

Tests

149 passing, mypy --strict clean, make publish-check clean.

Coming in v1.0

Phase 3 — Chrome MV3 extension surface for one-click dashboard / deep-dive launch. The local FastAPI daemon was already shipped in 0.1.0; the extension UI hasn't been built yet.

🤖 Generated with Claude Code

v0.1.0 — first publishable release

First publishable release of yc-ai-pulse.

End-to-end pipeline that pulls the latest YC batch, classifies it with a Sonnet-class model under strict anti-hallucination guards, and renders a single-file HTML dashboard with row-level drill-downs.

Highlights from the W26 dogfood run

• 63.3% coverage of the 196-company batch. 132 in upstream feed, 124 Tier A+B, 8 dropped (named in the register), 4 with dead websites.
• 95% high-confidence rate on the LLM enrichment (118 of 124).
• 0 schema-validation failures, 0 hallucinated source URLs.
• Top finding: 58% of high-confidence W26 companies build agents — defensible with row-level evidence in the dashboard heatmap.

A real run is checked in as a working example: examples/output/dashboard-w26-pr4-2026-05-01.html. Quality writeup at docs/QUALITY_REPORT_W26.md.

Quickstart

pipx install yc-ai-pulse
ycai run-coverage --batch winter-2026 --yc-official-count 196 --enrich

--enrich uses your Claude Max subscription via claude-agent-sdk by default (~6 min on W26, ~free). Pay-per-token via ANTHROPIC_API_KEY.

What's in 0.1.0

• yc-oss/api scraper with hard-fail on upstream unreachable (no companies?... scraping per robots.txt)
• PII sanitizer (idempotent, runs before disk and before LLM)
• Async link verifier with publish-gate enforcement on cited URLs
• Coverage probe (Tier A/B/C) with a dropped register that names every excluded company
• LLM enrichment with anti-hallucination Layer 1 — pydantic schema, source-URL grounding, two-pass cross-check
• Dashboard with capability×industry heatmap, tech-stack and OSS-posture breakdowns, confidence breakdown — all with drill-downs
• ycai resume <run-dir> for interrupted runs
• ycai dashboard <run-dir> for re-rendering at zero LLM cost
• 10 hallucination-trap fixtures locked in as regression tests
• 103 tests, mypy --strict clean, secret-scan clean

Coming in 0.2

• B007: depth=1 website crawl to recover tech_stack and oss_posture from unknown (the biggest signal lever)
• VC-style .pptx deck and .docx memo (Phase 2 of the project plan)
• Chrome extension surface (Phase 3)

PyPI Trusted Publishing setup

The release workflow is configured to publish to PyPI via Trusted Publishing. Setup is one-time on PyPI: https://pypi.org/manage/account/publishing/ — workflow release.yml, environment pypi, repo RyanAlberts/yc-ai-pulse. Until configured, the publish job fails but the GitHub release still has the wheels attached.

🤖 Generated with Claude Code