Projects

  1. AutoHarnessBench landing page and overall leaderboard
    AutoHarnessBench: AI R&D at the Harness Layer

    A long-horizon benchmark evaluating harness-driven agent self-improvement.

  2. Privileged self-distillation: rollout, construct hint, distill
    Privileged Self-Distillation: Post-Training from Corrected Rollouts

    A post-training method distilling privileged corrections to failed rollouts.

  3. Meta-Reward: evaluator harness optimized by meta-reward, feeding a frozen LLM judge
    Meta-Reward: Reward Modeling as Harness Optimization

    Framing reward modeling as a harness optimization problem.

  4. 67%
    80%
    87%
    Meta-Agent: Continual Learning for Agents

    An open-source framework for agent continual learning.

  5. Canvas Coworker — AI coworker for knowledge work
    Canvas Coworker: AI for Knowledge Work

    A production coding agent for long-horizon GTM and knowledge-work tasks with 100+ integrations.

  6. Goldfish long-video retrieval framework architecture
    Goldfish: Vision-Language Understanding of Arbitrarily Long Videos

    A SOTA vision-language model for question answering on hour-long videos.

  7. SlowFormer energy attack on vision transformers
    SlowFormer: Adversarial Robustness for Efficient Vision Transformers

    Adversarial robustness for vision transformers.

  8. Continual self-supervised learning with a frozen teacher preserving old representation geometry
    Continual Self-Supervised Learning with Knowledge Distillation

    A Twitch video embedding model that continually learns from video streams.

  9. Do Function Vectors Factor Task and Distribution?

    An interpretability study of how LLMs represent tasks during in-context learning.

  10. A human and an LLM choose different quantifiers for the same quantity
    Can LLMs Understand Quantifiers Like Humans?

    A study comparing human and LLM quantifier choices using Rational Speech Act models.

  11. MiniGPT4-Video architecture: interleaved visual-textual tokens through ViT and LLM
    MiniGPT4-Video: Advancing Multimodal LLMs for Video Understanding

    A multimodal LLM that interleaves visual and textual tokens for video understanding.

  12. DashGPT landing page — Data to Dashboard magic
    DashGPT: Spreadsheets to Interactive Dashboards

    A system that turns spreadsheets into interactive dashboards.

  13. Mitigating Negative Transfer in Multi-Task Learning with EMA Loss Weighting

    EMA-based loss weighting that reduces negative transfer in multi-task learning.

  14. Deep Learning-Based Autism Spectrum Disorder
    Detection Using Emotion Features From Video

    Detecting autism spectrum disorder from emotion features in video recordings.

  15. TikTokFER identity resolution across videos and duplicate removal
    TikTok for Good: Creating a Diverse Emotion Expression Database

    A diverse emotion-expression dataset collected from TikTok videos.

  16. Komma group-meeting scheduling landing page
    Komma: Group Scheduling for Sales Teams

    A group-meeting scheduling product for sales teams. Co-founder.