Edward Jex

Solve Intelligence (YC S23)

• Rewrote the core agent loop as its own microservice, out of the API. Agents now run in the background for hours, survive deployments, and scale to thousands at once. Results stream over Redis and server-sent events to the frontend, where they trigger actions; a reconnecting client resumes from a cursor with no gaps or duplicates.
• Took Charts from prototype to a production product used by thousands of customers, working end to end: the agent loop and AI tools, a planner delegating to subagents, through to the rich-text, cell-based editor. Tens of hours of attorney work now take minutes.
• Scaled agents from short, bounded tasks to long, open-ended ones: lazy document loading and parallel OCR for time to first token, and hunting down GIL-bound work to fit many agents on one container.
• Shipped across the stack: Postgres row-level security, live sharing, and patent language and jurisdiction conversion.
• Trained a 227,746-parameter (834 KB) model that labels each OCR character as struck or underlined, at 3 ms a line on two CPU threads, replacing image tokens at a sixth of the context cost. Took it apart with a pre-registered, 76-run retraining ablation: a placebo set the noise floor at a quarter of a point, and removing a part and retraining cost far less than switching it off (3.7 points against 94).

MEng thesis, supervised by Dr Farshad Arvin

Asked how little energy a useful detector can spend. Treated detection as one heatmap peak per animal rather than boxes, trained five hundred architectures across a GPU cluster and measured each on a Hailo-8L, because architecture explained only 15% of the variance in energy on the accelerator (80% on CPU). The best has three layers and 12,387 weights, reaching 0.948 AP at 2.2 µJ a frame, 680× less than YOLO11-nano. It runs in the browser on hand-written WGSL compute shaders, with every layer's feature maps inspectable.

write-up · code

A class-conditional diffusion model with classifier-free guidance in under a million parameters, reaching FID 18.1 on CIFAR-100, and a 100k-parameter MobileNetV3 variant with grouped pointwise convolutions.

Compared four model-free methods, then extended TD-MPC2 to solve BipedalWalker in under 100 episodes.

A UVCGAN2-style ViT-UNet CycleGAN, extended with MoCycle-GAN motion losses from RAFT optical flow, trained on an A100.