02 / story

Perception, policy,
and the real world.

I grew up in India taking apart the question that still drives me: how does a machine move through the world without being told every step? That curiosity ran from mechanical engineering into how machines perceive, reason, and decide from data.

Everything since has circled the same place: perception, policy, and real-world deployment. I am drawn to systems that have to behave reliably when conditions go off-script, which is usually the part a demo quietly leaves out.

I would rather understand something than offload it.

I build the hard parts from scratch because rebuilding a thing teaches you what running it never will. I built Feynman-Loop to hold myself to that: it makes me explain what I learn, and what AI writes for me, instead of letting the understanding slide. It is the opposite of letting a model do the thinking.

Understand, don't offload

Running a model and being able to rebuild it are different kinds of knowing. I want the second one.

From scratch first

DDPM and DDIM, a Vision Transformer, the transformer itself from Karpathy's lectures. The hard parts, by hand.

Reliable when it counts

I like the part most demos skip: behaving correctly when the world stops cooperating.

the path

  1. Grew up · India

    Gully cricket, and an early fixation on how things move. That turned into mechanical engineering, then into the software that makes machines decide for themselves.

  2. 2020 – 24 · Singapore · NTU

    B.E. Mechanical Engineering, minor in Computing & Data Analysis. My final-year project used neural networks in MATLAB to monitor electric-drivetrain health for mobile robots, and earned a Final Year Project Research Award.

  3. 2023 · Singapore · Venti

    Autonomous-vehicle test engineer. Tested autonomous prime movers at a Singapore port: lane detection, curb detection, and path-change validation with RViz simulation and LiDAR analysis. Watching a stack meet the real world is what made me take edge cases seriously.

  4. 2024 – 25 · Singapore · UOB

    Site Reliability Engineering on production banking infrastructure. A year living with the discipline reliable systems demand, where the one rare failure is the one that matters. I picked up tennis here too.

  5. since 2025 · Silicon Valley · Northeastern

    MS in AI (GPA 3.92). Taught Foundations of AI as a TA, and founded Φ, a robotics group, with three classmates. There I fine-tune vision-language-action policies for an SO-ARM101 arm and train them with RL in simulation. My research asks whether diffusion planners beat deterministic ones in nuPlan closed-loop. Under standard imitation they do not, because the diffusion policy collapses to a near-deterministic one. I spent most of that project trying to prove myself wrong, and wrote up the null with the diagnostics that explain it.

Next: a Spring 2027 internship where perception, planning, and deployment meet. Say hello →