Home›Companies›Finic
Finic

Frontier AI to stop fraud, abuse, and financial crimes.

Finic helps teams stop fraud, abuse, and financial crimes with system on models that make accurate decisions in fractions of a second, and system two agents that automate investigations end-to-end.
Active Founders
Jason Fan
Jason Fan
Founder
Ex security & product @ Robinhood. University of Toronto Computer Science.
Founder @ Finic (YC W23). Caltech class of 2020, former software engineer at Robinhood.
Company Launches
Grayson: A universal decision model for fraud, risk, and financial crimes
See original launch post

Hey everyone đź‘‹ I'm Jason, CEO of Finic. Today we're launching Grayson, the first universal decision model trained to catch fraud, abuse, and financial crimes in real time.

TL;DR: Send Grayson any context (logins, transactions, case notes, call transcripts) and a set of questions. Get back a calibrated probability for every answer in under 150ms. On fraud and risk benchmarks it's as accurate as Claude Sonnet 5.5 at 1/100th the cost, and it beats Jev. Zero data retention by default. Request API access and get $25 in free credit.

uploaded image

❌ The problem

I joined Robinhood pre-IPO and built our fraud investigations platform, along with some of our most critical authentication controls. I saw the same problem over and over: many risk decisions need human-level judgment, but they come up so often that paying a human (or an AI agent) to make them is financially unworkable.

  • Should this new device require step-up authentication?
  • Should we allow this transaction, or block it?
  • Should this ticket go to fraud, or to customer support?
  • Does this call log show signs the customer is being scammed?
  • Is this fraud?

Today, teams answer these with some mix of:

  • Rules that are brittle and get gamed the week after they ship
  • Custom ML models that take months, need labeled data, and have to be retrained every time a question or data source changes
  • Frontier LLMs that take seconds, cost dollars per thousand calls, return prose you have to parse, and are confidently wrong
  • Humans or agents who can do it well, but not on every login, transaction, and ticket

âś… The solution: Grayson

Grayson is a decision model: you can't chat with it. You give it context and typed questions, and it returns only probabilities, so a rules engine, router, or review queue can use them directly.

// Request (abridged)
{
  "context": { "known_devices": [...], "events": [...] },
  "questions": [
    { "name": "is_ato", "type": "noul", "instructions": "Are there signs of an ATO?" },
    { "name": "top_device_family", "type": "choice",
      "instructions": "Which device family do they use the most?",
      "options": ["apple", "android", "other"] }
  ]
}

// Response
{
  "is_ato": { "value": true, "probability": 0.999 },
  "top_device_family": { "value": "apple",
    "probabilities": { "apple": 0.999, "android": 0.001, "other": 0.000 } }
}

uploaded image

  • Universal: any context, structured or unstructured, and any question: yes/no, multiple choice, or a score. No retraining when your data or processes change.
  • Calibrated: when Grayson says 0.8, it's right about 8 times in 10. So you can set thresholds, auto-approve the clear calls, send the close ones to review, and move the threshold later without retraining.
  • Fast and cheap: under 150ms of model time for typical requests, at $0.035 per million input tokens (about $0.11 per 1,000 requests on our benchmarks).
  • Private: zero data retention by default. We never store the context, questions, or answers you send.

Grayson is trained with Reinforcement Learning for Calibrated Decisions (RLCD). Credit to the TypeSafe team behind Jev for pioneering RLCD and the "System One" model. Grayson takes that idea and specializes it for fraud, risk, and trust & safety.

📊 Benchmarks

We ran Grayson against Jev, Claude Opus 5.5, Claude Sonnet 5.5, GPT-6.1 Sol, Kimi K3, and Laya on six fraud and risk datasets: four public benchmarks (LendingClub, customs fraud, fake job postings, scam/phishing email) and two of our own held-out evaluation sets.

uploaded image

uploaded image

Grayson has the best log loss and the highest % right of any model, which matters most for a model whose job is to output probabilities you can threshold on. On the "is it fraud?" questions with known outcomes, it has the best log loss of every model tested:

uploaded image

And on general business decisions (JevBench and EIKOS), Grayson is level with Jev (0.971 vs. 0.974 AUC-PR), while frontier models score 0.988–0.998 at 100–500x the cost.

Full methodology, datasets, and charts are in the announcement post.

🙏 The ask

  • If you run fraud, risk, or trust & safety at a fintech, bank, lender, or marketplace: request API access. Every account gets $25 in free credit. We're in early beta (GPUs are still hard to get!) and are rolling out access gradually.
  • Intros: we'd love to meet heads of fraud, risk, and T&S, and the engineers who own rules engines and decisioning platforms.
  • Questions or ideas for what Grayson should answer? Email me at jason@finic.ai.
Previous Launches
Helping fintechs increase fraud ops capacity, improve accuracy, and resolve cases 80x faster.
Get acquisition offers in just 3 months. No exclusivity, cancel anytime
YC Photos
Finic
Founded:2025
Batch:Winter 2023
Team Size:4
Status:
Active
Location:San Francisco