An unreleased prototype Raj built to support long browser workflows. It supplies compressed whole-page context, tracks appearing and disappearing elements, and detects unsuccessful clicks. Raj discussed open-source, API, website, and plugin delivery as prospective options rather than completed releases.
A public Hugging Face profile that explicitly links to the kushan-sarvam GitHub account. The retained snapshot showed no public models or datasets.
huggingface.co
Bio, Work & Ideas
Kushan Raj
Conference affiliation: Sarvam AI
On this page
Kushan Raj is an AI engineer who helped build Sarvam Samvaad, a platform for multilingual conversational agents, and develops infrastructure for browser agents. His work addresses the practical distance between a model’s abilities and a useful product: responding naturally over a phone call, understanding a web page, and recognizing when an attempted action has failed.
From medical imaging to conversational agents
Raj studied bioengineering and complex systems and dynamics at IIT Madras. His thesis investigated diffusion models for medical imaging, examining how generated images could help address scarce and variable data. Working with adviser Srinivas Chakravarthy, he explored denoising diffusion probabilistic models and techniques including latent diffusion, ControlNet, and classifier-free guidance. This earlier work placed generative models in a concrete application where producing plausible output was only part of the problem; the intended use was to support medical research and image analysis.
At Sarvam, Raj moved into real-time conversational systems for Indian languages. He helped develop Samvaad over the year preceding its launch, with a particular enthusiasm for getting agents into use quickly and improving them through customer interactions. Language and voice could be configured for different callers, while conversation insights helped teams revise agent behavior.
The product brings together speech recognition, a language model, and speech synthesis in a real-time loop. Its shared agent design—prompts, tools, knowledge, and variables—supports both spoken and written conversations across channels such as telephony and WhatsApp. Sarvam subsequently renamed the platform Voice Agents. Raj’s contribution was to building the conversational product; the broader platform and its underlying models are collaborative company work.
Giving browser agents useful context
Raj’s browser-agent work extends his interest in making AI useful through the surrounding software. His browser runtime prototype focuses on the information a model receives and the feedback it gets while executing a sequence of actions. Slow, uncertain interactions are an adoption problem: automation has little appeal when completing the task manually is easier.
Two ideas define his approach:
Compact whole-page context: The runtime supplies a Markdown representation covering the page, which can accompany a screenshot. This gives the agent information beyond the current viewport without feeding it the entire document object model. Raj’s goal is to make relevant page structure available at a manageable token cost.
Feedback about action outcomes: The runtime tracks appearing and disappearing elements, removed obstructions, and clicks that did not take effect. An agent can then revise its next action using what actually changed, rather than assuming that issuing a command completed the step.
As of June 2026, Raj described open-source distribution and an API accepting a URL and user intent as possible next steps for the prototype. His ambition is practical: browser agents that are fast and dependable enough for ordinary online work, with improvements coming from better page representations and execution feedback as well as model capability.
Kushan Raj argues that browser agents often waste capable models on fragmented page views and ambiguous actions. His prototype instead gives the model compact whole-page context and explicit feedback about what changed after each interaction.
A capable model can still waste time when its browser interface supplies fragmented observations and weak confirmation that actions succeeded.
For the demonstrated page, Raj compares a roughly 20,000-token DOM with an approximately 1,100-token screenshot and an approximately 1,800-token whole-page Markdown representation.
State deltas—what appeared, disappeared, became unblocked, or failed to change—give the model a clearer recovery signal than repeated screenshots alone.
The examples support an interface-design hypothesis, not a general performance result: the talk provides no controlled success-rate, latency, cost, safety, or information-loss evaluation.
▶ Watch ↗