Saptadip Saha

Background:

I am a senior undergraduate in Computer Science and Engineering. Rather than list topics, I describe the causes I am working on or intend to work on.

As models scale, oversight has to scale with them. This motivates my interest in alignment, interpretability, red-teaming, and capability evaluations. Beyond scaling oversight, I want to apply AI where it is needed, not forced, to interdisciplinary problems such as astronomy, synthetic biology, automated formal proofs, and cybersecurity. My aim is not to become a specialist in these fields, but to consult extensively with domain stakeholders to understand the problem and its requirements, and then apply AI accordingly. This requires work on generative models, particularly multimodal foundation models, and equally on the agent harness: its components and primitives, agent containment and sandboxing, continually improving (recursive) harnesses, and agent swarm. Much of this borrows concepts from operating systems, system design, software engineering, and computer and network security, which I find more interesting to work on given my background.

Working Style and Preferences:

In one word, it would be "Agentic". I am always eager to own my work, take responsibility, and collaborate extensively. I am always ready to ideate, brainstorm, and iterate. Rather than being completely assured of the roadmap before starting, I prefer an agile approach. While working on a project, I equally emphasise following standard software engineering design patterns and principles to ensure fast debugging, maintainability, scalability, and transparency, right at the codebase level. Therefore, I am not a great fan of delegating a high-level idea and project plan to a swarm of coding agents and only validating the end result. Frankly, I don't like vibe-coding; to me, a more appropriate term is Accelerated Coding. I emphasise owning the solution, and I see AI agents as a means of writing code faster (as humans have limited typing speed) and of brainstorming and validating ideas by generating many edge cases and counterfactuals. While working with others, their preferences may differ, but I believe a team is better off following the "Whiteboard Defense" practice: understand the work well enough to explain it, defend it, and challenge it.