scrape-company

name	scrape-company
description	Use when researching a Prague AI company to build its catalog entry. Defines what data to collect, where to look, and the exact JSON output schema.

Scrape Company

Collect structured data on a Prague AI/ML company for the job-seeker catalog. One company per invocation. Output is a single JSON file saved to data/companies/{id}.json.

Input

The agent receives a company record from data/companies_index.json:

{"id": "rossum", "name": "Rossum", "category": "company", "cnaip_url": "https://www.cnaip.cz/cs/entity/rossum", "website_url": "https://rossum.ai"}

Scraping Strategy (ordered by reliability)

Source 1 — CNAIP page (ALWAYS start here if available)

CNAIP pages are consistent and scrapable. Navigate to cnaip_url, take snapshot. Extract: description, tags, team members, links to company website.

Source 2 — Company website

Use website_url from index (do NOT waste time searching for it). Visit in order:

1. Homepage — tagline, what they do
2. About/Company page — founding year, history, team size, offices
3. Careers page — open roles, tech stack signals

If the site blocks you (captcha, anti-bot, redirect): STOP. Do not retry. Move to Source 3.

Source 3 — Search snippets (fallback)

Search for "{company_name}" Prague AI and extract from snippets. Good sources:

• tech.eu, TheRecursive, AIN.Capital articles
• Tracxn, Owler summaries (often in snippets)
• LinkedIn company page snippets

Source 4 — Funding (best effort only)

Search "{company_name}" funding raised series site:tech.eu OR site:therecursive.com OR site:ain.capital. Do NOT attempt Crunchbase or PitchBook (they block automated access).

Source 5 — Glassdoor

Search "{company_name}" glassdoor rating reviews. Extract:

• Overall rating (1.0-5.0)
• Number of reviews
• "Recommend to a friend" percentage (if visible in snippet)

Do NOT navigate to glassdoor.com directly (login wall). Use search snippets only.

Source 6 — Revenue (best effort only)

Search "{company_name}" revenue ARR. Accept whatever snippets return. Do NOT spend more than 2 searches on this — revenue data for private Czech companies is rarely public.

Enum Values (MANDATORY — use EXACTLY these values)

Copy-paste from this list. Do not improvise.

business_type:      b2b_saas | b2c_saas | consulting | product | research | enterprise_vendor | other
company_type:       startup | scaleup | enterprise | acquired | research_lab | nonprofit
ai_focus:           nlp | cv | ml_ops | generative_ai | robotics | data_engineering | fraud_detection | recommendation | speech | automation | document_processing | other
employees_range:    1-10 | 11-50 | 51-200 | 201-500 | 501-1000 | 1001-5000 | 5000+
business_model:     subscription_saas | usage_based | consulting_services | licensing | marketplace | freemium | other
hiring_signals:     actively_hiring | few_roles | not_hiring | unknown
data_quality:       high | medium | low
risk_level:         high | medium | low | not_applicable

Rules:

• employees_range — round to nearest bucket (e.g. 70 employees → "51-200", NOT "51-100")
• business_type — use "b2b_saas" not "B2B", "B2B SaaS", etc.
• data_quality — high = 3+ sources, medium = 1-2 sources, low = estimated/inferred
• If you're unsure which enum value to use, pick the closest match or use "other"

JSON Output Schema

Save to data/companies/{id}.json:

{
  "id": "rossum",
  "name": "Rossum",
  "category": "company",
  "scraped_at": "2026-03-28T12:00:00Z",

  "overview": {
    "description": "AI document processing for transactional workflows",
    "description_long": "Rossum builds AI agents that read documents, capture and validate data, send emails, handle approvals, and write to ERPs. Proprietary transactional LLM supporting 276 languages.",
    "tagline": "Offload paperwork to AI agents",
    "website": "https://rossum.ai",
    "cnaip_url": "https://www.cnaip.cz/cs/entity/rossum"
  },

  "classification": {
    "business_type": "b2b_saas",
    "ai_focus": ["document_processing", "nlp", "cv"],
    "industries_served": ["finance", "logistics", "retail", "manufacturing"],
    "company_type": "scaleup",
    "is_ai_native": true
  },

  "history": {
    "founded_year": 2017,
    "founders": ["Tomáš Gogar", "Petr Baudiš", "Tomáš Tunys"],
    "origin_story": "Founded by 3 Czech PhD students in AI",
    "key_milestones": [
      {"year": 2017, "event": "Founded in Prague"},
      {"year": 2024, "event": "Named in Sifted B2B SaaS Rising 100"}
    ]
  },

  "size": {
    "employees_range": "201-500",
    "employees_exact": null,
    "employee_source": "company_website",
    "office_count": 5
  },

  "offices": [
    {"city": "Prague", "country": "CZ", "is_hq": true},
    {"city": "London", "country": "GB", "is_hq": false}
  ],

  "financials": {
    "publicly_traded": false,
    "stock_ticker": null,
    "stock_exchange": null,
    "parent_company": null,
    "funding_total_usd": null,
    "funding_rounds": [
      {"date": "2021-06", "type": "Series A", "amount_usd": 10000000, "lead_investor": "Some VC"}
    ],
    "investors": ["Investor A", "Investor B"],
    "revenue_estimate_usd": null,
    "revenue_source": null,
    "business_model": "subscription_saas"
  },

  "team": {
    "leadership": [
      {"name": "Tomáš Gogar", "role": "CEO", "source": "company_website"}
    ],
    "notable_people": []
  },

  "careers": {
    "careers_url": "https://rossum.ai/careers/",
    "open_roles_count": null,
    "open_roles": [
      {"title": "ML Engineer", "department": "engineering", "location": "Prague"}
    ],
    "tech_stack": ["python", "typescript", "aws"],
    "hiring_signals": "actively_hiring"
  },

  "glassdoor": {
    "rating": 4.2,
    "review_count": 35,
    "recommend_pct": 85,
    "glassdoor_url": "https://www.glassdoor.com/Overview/Working-at-Rossum-EI_IE..."
  },

  "recognition": {
    "awards": ["IDC Leader 2023", "G2 Leader Fall 2024"],
    "notable_clients": ["Siemens", "Bosch", "Wolt", "Adyen"],
    "customer_count": 450,
    "partnerships": []
  },

  "risk_assessment": {
    "agi_displacement": "low",
    "agi_displacement_rationale": "Proprietary transactional LLM trained on domain-specific document data across 276 languages. Deep vertical moat in document processing with enterprise integrations that foundation models cannot easily replicate.",
    "market_risk": "medium",
    "market_risk_rationale": "IDP market is growing but increasingly competitive with Microsoft, Google, and AWS offering native document AI. Rossum differentiates on workflow automation beyond pure extraction.",
    "funding_risk": "low",
    "funding_risk_rationale": "Well-funded scaleup with 450+ paying enterprise customers and strong revenue signals (Sifted Rising 100).",
    "overall_risk": "low",
    "overall_risk_rationale": "Strong proprietary tech moat, established enterprise customer base, and vertical focus in regulated industries provide resilience against AGI commoditization."
  },

  "metadata": {
    "data_quality": "high",
    "sources_used": ["company_website", "cnaip", "tech.eu"],
    "fields_missing": ["funding_total_usd", "revenue_estimate_usd"],
    "needs_review": false,
    "notes": ""
  }
}

Risk Assessment Guidelines

Every company MUST have a risk_assessment section. Assess based on what you learned during scraping. Use risk_level enum: high | medium | low | not_applicable.

AGI Displacement Risk

How likely foundation models (GPT, Claude, Gemini) will commoditize this company's core value proposition.

• high: Thin wrapper around LLM APIs, generic chatbot/automation, no proprietary data or tech. Easy to replicate with a prompt.
• medium: Vertical SaaS with some domain expertise, but core AI capability is becoming commoditized. Has customer lock-in that buys time.
• low: Deep proprietary data moat, infrastructure layer that AGI still needs, regulated domain requiring compliance/certification, or hardware/robotics.
• not_applicable: Research groups, nonprofits, government initiatives.

Market Risk

Is the market real, sustainable, and big enough?

• high: Hype-driven niche, no clear PMF, tiny addressable market, or dominated by well-funded incumbents.
• medium: Real market but competitive, or dependent on a single trend/platform.
• low: Established market with proven demand and paying customers.
• not_applicable: Research groups.

Funding Risk

Can the company survive long enough to reach sustainability?

• high: No known funding, pre-revenue, small team with no runway signals.
• medium: Some funding but early stage, or bootstrapped with revenue but unclear margins.
• low: Well-funded (Series A+), profitable, or backed by large parent company.
• not_applicable: Research groups (grant-funded), acquired companies, enterprise subsidiaries.

Overall Risk

Weighted judgment across all factors. A company with high AGI displacement risk but low market and funding risk might be medium overall.

Rationale is required for every risk rating — 1-2 sentences explaining why, referencing specific facts from your research (product type, funding, customer base, tech moat).

Null Handling

Use null for unknown fields. Never invent data. If estimated, note source in metadata.sources_used and set data_quality accordingly.

Research Group Adaptation

For category: "research_group":

• classification.business_type = "research"
• classification.company_type = "research_lab"
• financials — focus on grants, not funding rounds
• careers — PhD positions, postdocs, research assistants
• team.leadership — PI / group lead
• recognition — top publications, h-index of PI

Agent Dispatch Pattern

To scrape multiple companies, dispatch parallel agents (max 5-10 at a time):

for each batch of companies from companies_index.json where status == "pending":
  dispatch agents, each with:
    - company record from index
    - this skill loaded
  on completion:
    - save {id}.json to data/companies/
    - run validate_company.py on the output
    - update companies_index.json status to "done"

Post-Write Validation

After writing JSON, the orchestrator MUST run python3 scripts/validate_company.py data/companies/{id}.json to catch enum mismatches and missing required fields. If validation fails, fix the file before marking the company as done.