• 1 credit = 1 record returned. Empty results are free.
• Sizing trick: max_results: 1 returns the total match count, so a whole segment is sized for 1 credit.
• Key draws from OneGTM's shared 13.5M pool, so our own 10,000 cap has to be self-enforced.

• blitz.py logs every call (endpoint, request, records, balance) to credit_log.csv.
• Refuses any pull over 500 records without explicit approval.
• Rule for every phase: test → verify by hand → scale, always with a cap.

• Missing Ray Summit / Part 2 file and the "two shared sections".
• Closed-won list beyond the logo wall.
• Segments already ruled out, geography, dollar price per credit.

• Looked each up in Blitz (15 credits).
• Pattern: small vertical software companies (10–200 people) whose product eats messy paperwork. VLM Run sits inside their product.
• Industry labels are all over the place (4 labels for 4 customers), so description keywords have to do the work.
• Peak Health: the guessed domain matched the wrong company. Excluded. Lesson: verify domains from the source.

• Ranked by evidence: healthcare ops = 3 of 4 customers (high). Construction = CivilGrid (medium). The rest = tests.
• Out: doc-AI competitors, integration partners, services shops, >1,000 employees, the hospitals/GCs themselves.

• v1 loose searches matched ~12,000 companies, about 70% junk.
• Biggest fix: software-only industries. Construction went from 0 good to 6 of 10.
• Multi-word phrases are AND-of-words ("physical AI" matched a physical-therapy EHR).
• Competitors look exactly like prospects → blocklist of 29 names.

• S7 Video AI: 3 of 15 "maybe", mostly generative video and ad tech.
• J2 CV hiring: teams hiring core CV engineers are building their own (build, not buy). Kept as a signal idea instead.
• J1 document-AI hiring: low volume, highest intent → saved for signals.

• Changing a filter restarts results at page 1, so you pay again for rows you already have.
• Junk here was cheap to remove locally, so we resumed instead of re-filtering.
• Finds: Avallon AI, Heron Data, Two Dots, KlearNow, Alvys, Humata.

• Construction: "reality capture" pulled motion-capture companies. Re-filtered.
• Robotics: half were integrators or hardware. Tilted to robot learning, autonomy and teleop.
• Re-filter only when it saves more junk than it re-buys (~15 dupes in robotics).

• 35 companies checked on their public job boards (Greenhouse, Lever, Ashby, Workable). Free.
• 104 checked with Blitz jobs data. Empty results are free, so it cost 28 credits.
• Result: 29 companies hiring ML/AI engineers right now.

• How many engineers each company has, and what share of staff that is. Tells us whether they can integrate an API.
• 129 credits for the strongest 130 in this phase, then the other 64 at the start of Phase 4, so all 194 are covered.
• Blitz's docs say this data is free. It actually costs 1 credit per company. We caught that on the first few calls.

• Visual core: yes 30, partly 12.
• Messy input: fax, handwriting, scans, phone photos.
• Modality: docs 5 · images 10 · video 15 (VLM Run's edge).
• Extracting today: OCR, IDP or GPT/Claude stack = existing budget.

• Sensitive: HIPAA/PII 6 + VPC/on-prem/ZDR 4. SOC 2 alone doesn't count.
• Timing: ML/AI hiring 7 + "launching soon" 3.
• Buyable: 10–300 people 6 + ≥5 engineers 4.

A perfect buyer with no visual workload can't score above ~42.

Faxes, referrals, ACORD forms, pay stubs. Lead with accuracy on messy docs + ZDR, BAA, VPC. Proof: Valerie Health, Scan.com.

Drivers photograph BOLs and PODs; jobsite photos and plans. Lead with docs + photos in one API. Proof: CivilGrid.

Footage at scale that needs searching. Buyer is the ML/data team. No customer proof yet.

• Funnel: 1,000 → 349 (35%) → 100 (10%).
• Inference: the other ~1,300 hold another 100–130 top-grade accounts.
• Loose universe ~12,000 at ~30% fit → low thousands of real fits.

• Batch 2 alone: ~16 per top-100 account.
• Next ~1,300 matches ≈ 2,600 credits: above our remaining working cap, so not run.
• Full loose universe ≈ 20,000+ credits (inference). Needs OneGTM's budget and $ per credit.

• Keyword splitting · filter changes restart pages · docs wrong on cost · shared pool
• Fake evidence · search labels ≠ what they do · unreadable sites / LLM drift
• Stale signals · best segments run dry → grow by geography and hiring

• Company: ML/doc-AI hiring, job post naming Textract/Tesseract, new intake product page, compliance page, funding, marketplace listing, public LLM-vision code, repo activity, VLM Run signup.
• People: new CTO/Head of AI (90 days), customer champion moved, OCR-pain post, event speaker, engaged with VLM Run posts.
• Strong signal on a top-100 account → signal email within 48h. One signal email per account per 30 days.

• Blitz people search → Blitz waterfall ranks decision maker > influencer > champion.
• Email-finder waterfall → MillionVerifier: valid → sendable.csv.
• Catch-alls → BounceBan → deliverable joins the same file. Rest → LinkedIn only.

• 25–55 words, one fact each, CTA offers to run their samples and send the JSON.
• All 12 emails pass the house spam scan.
• LinkedIn: lowest-scored accounts first. Founder + growth account connect the same day, then 30–60 days of silence while weekly posts warm them up. Only the employee account messages.

• Inflooens: Claude Sonnet 4 for PDFs, GPT-4o-mini for images.
• Coral: e-fax extraction + hiring ML engineers.
• Zeitro: borrower phone photos → 1003.
• Alvys: drivers snap BOL/POD photos.
• Blitzz: OCR during live video sessions.
• At 1,000: the same website read writes the line from a verbatim quote. No quote → segment fallback.