Layanne El Assaad
Most of my work is on reliability infrastructure for machine learning and agent systems. M.S. in Computer Science at Columbia.
Currently building tokenslash.ai to cut AI agent inference costs. If your agents cost too much, get in touch.
Projects
- Agent infrastructure
Scout
Scout is a macOS platform for running and managing AI agents locally. Agents run in a sandbox, act only through permissions the user grants, and share a local knowledge graph of the user's files and email. All user data stays on device.
- Inference performance
Profiling ESMFold
ESMFold predicts protein structure from a single sequence, but its forward pass grows quadratically with sequence length. This project profiles it on an H100 to find which stages dominate on long proteins and how much graph compilation recovers.
- Multi-agent systems
ESCAPER
ESCAPER is an escape-room environment for studying trust calibration under partial observability, where agent teams solve rooms together while a hidden malicious teammate feeds them misinformation.
- Distributed training
Pipeline parallelism schedules
A benchmark of the three main pipeline parallelism schedules, GPipe, 1F1B and interleaved 1F1B, across model shapes and process counts. It measures throughput, scaling efficiency and activation memory for each, and shows where one schedule overtakes another.