pitchgrill

Get grilled before the VCs do.

pitchgrill reads your pitch deck and tells you where a skeptical investor will dock you points: the red flags that kill deals in diligence, the questions that will catch you flat-footed in the room, and the data-room documents you don't have yet.

It is not a scorer. It does not predict your odds of success (early-stage base rates are too low for that number to mean anything). It does one useful thing: surface the losing moves before an investor does.

中文说明见文末。

Use it in your browser — no install

→ layton2617.github.io/pitchgrill

Pick your stage and sector and you immediately see the red flags, the questions you'll be asked, and the data-room documents you'll need. No signup, no API key, nothing to install. Upload your deck PDF (or paste text) to sharpen the ranking. With an Anthropic API key you also get a line-by-line read of your specific deck — image-heavy PDFs included — run entirely in your browser.

Want visitors to run the AI analysis without their own key? Deploy the optional Cloudflare Worker in serverless/ — it holds your key server-side, rate-limited so a public endpoint can't run up your bill.

The rest of this README is for running it locally or from the command line.

Why not just ask ChatGPT

Paste your deck into a chatbot and you get polite, generic encouragement. pitchgrill ships a structured knowledge base that pins down what investors actually look for, as checkable rules:

• Red flags with thresholds — not "customer concentration looks a bit high" but "single customer > 30% of revenue = major flag, > 50% = deal-killer."
• Stage-specific grilling — the questions a pre-seed founder gets are not the ones a Series A founder gets. Each comes with what a strong answer sounds like and the weak answer that gives you away.
• Data-room checklist — the documents diligence will ask for, split into must-have vs. nice-to-have.
• Deck lint — the narrative mistakes that get a deck rejected on sight.
• Sector benchmarks — healthy ranges and red lines per sector and metric.

Everything lives in plain YAML under kb/. You can read every threshold and its source, and edit any of it to match your own view.

Quickstart

pip install -r requirements.txt

# Plain text or PDF deck. --founder is an optional free-text note.
python check.py \
  --deck examples/sample_deck.md \
  --stage seed \
  --sector dev-tools \
  --wedge ai-dev-tools

Set ANTHROPIC_API_KEY and pitchgrill reads your specific deck and tells you which line triggers which red flag. Without a key it falls back to a checklist mode: the most relevant items for your stage and sector, ranked, with no per-deck analysis.

Example

For a seed-stage AI dev-tools deck where one customer is 81% of revenue and gross margin is 38%:

## Red flags

🔴 KILL  Revenue concentration in a single customer
  threshold: single customer > 50% of revenue
  evidence:  MegaCorp is $340K of $420K ARR (81%)

🔴 KILL  Gross margin far below sector norm
  threshold: SaaS healthy 70-80%+; services-heavy delivery is a flag
  evidence:  38% margin, plus 2 dedicated delivery engineers

## You will be asked

- "If MegaCorp churns tomorrow, are you still a seed company?"
- "38% margin with dedicated engineers — is this software or consulting?"
- "Walk me through the $50B TAM. Bottom-up, what can you actually reach?"

## Data-room gaps

- Revenue by customer / cohort
- Fully-diluted cap table (including the verbal advisor grant)
- Unit economics: CAC / LTV / payback / fully-loaded margin

Wedges

Three knowledge layers. general always applies; pass --wedge to stack a niche layer on top:

• general — applies to everyone.
• cross-border-cn — Chinese / cross-border teams raising USD (VIE structure, ODI, US-VC objections to a China-based team).
• ai-dev-tools — AI and developer-tools startups (token-cost margins, "why won't the model provider just build this", data-moat questions).

How it works

check.py                 entry point
src/pitchgrill/
  cli.py                 arg parsing, reads the deck (txt/pdf)
  kb.py                  loads kb/ yaml, filters by stage / sector / wedge
  engine.py              picks relevant items, builds a grounded prompt, calls the model (or falls back)
  report.py              renders the three sections
kb/                      the knowledge base (yaml)
scripts/                 maintenance: rebuild and source-check the kb
schema/kb_schema.md      the kb data shape
examples/                sample deck

The model only maps your deck onto the knowledge base. It does not invent judgment criteria beyond what's in kb/, which is why the output is auditable: every red flag points back to a rule and a source you can read.

Customize

The knowledge base is the product. Edit the YAML in kb/ to change a threshold, add a red flag, or tune the grilling for your own thesis. Contributions of new sectors and wedges are welcome.

To rebuild kb/ from a single JSON file, or to re-check that every cited source URL is still live:

python scripts/explode_kb.py kb.json              # JSON -> kb/*.yaml
python scripts/check_sources.py kb.json --strip   # drop dead source links
python scripts/build_web.py                        # rebuild docs/kb.json for the web page

License

MIT. See LICENSE.

---

中文

pitchgrill 吃你的 pitch deck,告诉你一个挑剔的投资人会在哪里扣分:DD 时会让流程死掉的红旗、会在会议室里问倒你的问题、你还缺的 data room 材料。

不是打分器,不预测成功率,只做失分项自检。区别于裸用 ChatGPT 的泛泛反馈,pitchgrill 带一个 kb/ 下可读可改的结构化知识库:带阈值的红旗、分阶段拷问、data room 清单、deck lint、赛道 benchmark。设了 ANTHROPIC_API_KEY 就针对你的 deck 逐条命中,没设则降级为清单模式。三个 wedge:general / cross-border-cn(中国出海)/ ai-dev-tools。