Whitney AIRequest access

Journal

Research and field notes on evaluations, post-training, and better model behavior.

  • A pixel-art wooden Trojan horse releasing small robot agents through an open castle gate at night.

    Case Study9 min read

    Post-Training a 27B Model to Outperform Frontier Red-Teamers by 3x

    By Jake Kang, Vatsal Bajaj, and Ranjan Goel

    Whitney AI took a 27B open-source model and trained it into a red-teamer that outperforms attackers many times its size, 3.1x DeepSeek V4 Flash, with refusal-to-attack down from 81% to 2%.

  • A pixel-art robot lounging at a desk with its feet up, holding a red telephone handset to its ear.

    Case Study7 min read

    Scaling Scam-Call Data Without Scaling Phone Lines

    By Vatsal Bajaj, Jake Kang, and Nick Mecklenburg

    How Whitney AI post-trained a 12B active-parameter generator that writes scam-call transcripts a frontier judge can barely tell from real ones, so Scam AI can build fraud detectors without running thousands of honeypot phone lines.

  • A pixel-art health portal displayed on a CRT monitor, with a targeting reticle over a booking button.

    Case Study7 min read

    Faster Click Prediction for Browser Agents at CloudCruise

    By Jake Kang

    How Whitney AI post-trained a 9B-parameter click-prediction model for CloudCruise's browser-automation builder agent, matching Gemini 3 Flash on accuracy while running 9x faster and 5x cheaper.

  • A pixel-art desk phone in a medical office, framed by a first-aid sign and a doctor portrait.

    Case Study6 min read

    Building Reliable Voice Agents at Harmonize Health

    By Jake Kang and Nick Mecklenburg

    How Whitney AI post-trained a 4B-parameter guardrail model that catches medical-advice violations in Harmonize Health's voice-agent calls before the sentence finishes, matching frontier quality while running 66% faster.