Portrait, photo 1 of 2

Essam Sleiman

I founded Canvas Labs (YC), where we build a coding agent and platform for long-horizon knowledge work.

Previously, I was a research engineer at Twitch and did research on RL, post-training, and continual learning as a grad student at Harvard, with papers at ECCV and CVPR. In my free time, I enjoy basketball, coffee, and hiking around SF.


Selected Work

Canvas Coworker — AI coworker for knowledge work
Canvas Labs Canvas Coworker A production coding agent for long-horizon GTM and knowledge work. Custom harness and platform with per-session cloud sandboxes, persistent files, scheduling, and 100+ integrations.
AutoHarnessBench landing page and complete overall leaderboard
Canvas Labs AutoHarnessBench Essam Sleiman, Mersad Abbasi, Ayush Chakravarthy, Karina Nguyen A long-horizon benchmark evaluating harness-driven agent self-improvement.
Privileged self-distillation: rollout, construct hint, distill
Canvas Labs Privileged Self-Distillation Essam Sleiman* A post-training method distilling privileged corrections to failed rollouts.
67%
→
80%
→
87%
Canvas Labs Meta-Agent An open-source framework for agent continual learning.
Goldfish long-video retrieval framework architecture
Research, Harvard Goldfish: Vision-Language Understanding of Arbitrarily Long Videos Essam Sleiman* (Contributor) A SOTA vision-language model for question answering on hour-long videos.
SlowFormer energy attack on vision transformers
Research, UC Davis SlowFormer: Adversarial Robustness for Efficient Vision Transformers Essam Sleiman*, K L Navaneet*, et al. Adversarial robustness for vision transformers.
Continual self-supervised learning with a frozen teacher preserving old representation geometry
Research Engineer at Amazon Science Continual Self-Supervised Learning with Knowledge Distillation Essam Sleiman*, Xiangbo Li, Saad Ali A Twitch video embedding model that continually learns from video streams.
Research, Harvard & MIT Do Function Vectors Factor Task and Distribution? Essam Sleiman*, Jacob Andreas An interpretability study of how LLMs represent tasks during in-context learning.

→ See all projects