Hi, my name is Sayak. I am interested in solving real-world challenges with machine learning. Currently, I live and work in the UK as a machine learnin...
While building compound agentic systems, some of the design dimensions that we consider optimizing are memory, context management (what the systems sees vs does not) and the accumulating history over longer horizons. These choices have direct downstream qualitative and quantitative effects like improving the agent's action outputs and the token/cache economics respectively. In this blog, I share…
I recently wrote some code for our internal agent harness and system architecture that I was particularly happy with. With the new Opus 4.6 release last week, I wanted to see if it could one-shot the same task. So decided to run this small experiment on the weekend afternoon. Asked Opus 4.6 and Codex 5.2 (no access to 5.3 at the time) the same assignment with the necessary modules and a prompt…
The project I have been working on for the past few months involves building enterprise agents to assist internal users with their workloads. For context, we are working with a consultancy that acts as an intermediary between educational institutions and clients. This blog discusses this experience, covering data pipelines, agent architecture and scaffolding, building metrics, evaluation loops,…
Optax is an optimization library for JAX which provides out-of-the-box apis for commonly used ml loss functions, optimizers and utilities like learning rate schedulers. In general, defining your own optimizers with Optax is quite simple. Use the optax.chain api to combine different transforms. Transforms are pure functions that act on model parameters to update their values according to the update…