Deep reinforcement learning
Actor–critic methods, off-policy learning, policy optimization, and sample efficiency.
RESEARCHER · BUILDER · EXPLORER
Hi, I’m
Shuvrajeet Das.Building agents
that learn to explore.
PhD student / Data Science & AI / IIT Madras
I study how intelligent agents learn through interaction. My work brings together reinforcement learning, generative models, and the engineering that turns an idea into an experiment.
COMPUTATION GRAPHCONCEPTUAL VIEW
→→
Start with a question.
Make the assumptions explicit. Understand the idea before writing the implementation.
INTUITION → IMPLEMENTATION● INTERACTIVE
THE PERSON BEHIND THE EXPERIMENTS
Curiosity, with a working implementation.
I’m Shuvrajeet, a PhD student in Data Science and Artificial Intelligence at IIT Madras, exploring how intelligent agents can learn through interaction.
I enjoy taking an idea from equations to a working implementation—and investigating what happens when the experiments disagree with the intuition.
My research focus is deep reinforcement learning and diffusion-based exploration. The question that keeps me coming back: how can learned generative priors help agents explore more effectively?
Alongside research, I’m building self-play agents, game-playing systems, and continuous-control experiments. I’m also learning probabilistic machine learning, optimization, and GPU programming.
See what I’m building →WHAT I THINK ABOUT
Six connected directions.
Learning. Decision-making. Exploration.
Diffusion models for RL
Learning action priors and investigating their role in exploration.
Exploration & uncertainty
Structured exploration, behavioral diversity, and probabilistic reasoning.
Self-play & planning
Game-playing agents, Monte Carlo tree search, and learning through competition.
Continuous control & robotics
Sequential decision-making in control tasks and embodied environments.
Efficient learning systems
Vectorized rollouts, GPU computation, and reproducible training pipelines.
02 / QUESTIONS WORTH EXPLORING
At the research frontier.
RESEARCH SPOTLIGHT
Can a learned prior
lead to better exploration?
I’m investigating diffusion models as a source of structured exploration for reinforcement learning: can a model propose useful exploratory actions while an RL agent learns how to refine and use them?
I’m interested in how those proposals interact with a learned actor, and how to evaluate their contribution through verified baselines and controlled experiments.
Read the research direction ↗Collected experience↓Learned action prior↓Structured explorationOngoing research · benefits not yet established
WHAT I’M BUILDING
From agents to robots.
Project descriptions come from my profile README. This website’s repository does not contain their implementations or experiment results.
INSIDE THE LABORATORY
Algorithm explorer.
Follow the implementation.
Understand the computation.
Reading repository inventory…
What does this catalog include?
Research projects mentioned in my profile are shown separately above. They are not counted as local implementations.
03 / THE COMPUTE LAYER
One idea. Many operations.
Hardware is part of the question.
Understand what
the machine is doing.
The README describes TensorFlow pipelines, vectorized environments, and multi-GPU training as engineering interests.
This diagram is conceptual. Device support, memory use, and speed depend on the actual implementation and must be checked per experiment.
NO LOCAL HARDWARE BENCHMARKSTensorFlow tensor operations
CPU
General-purpose compute
Control flow and numerical operations
GPU
Parallel compute
Supported operations across many elements
Conceptual execution paths · not a performance comparison
04 / EVIDENCE OVER ASSUMPTIONS
The experiment desk.
Experiment results · coming soon
No local results or plots were found in the inspected checkout. Research directions in the README are not measured outcomes.
05 / OPEN THE NOTEBOOK
A map of the work.
∴ TheUnsolvedDev / project inventory
READING PATH
Start at the source.
The README is the authoritative introduction to the research. Implementation-specific learning paths become available as source directories are added.
Open the README ↗HOW I BUILD
Less hype.
More experiments.
Let the ablations speak.
I work with TensorFlow-first model development, custom training loops, and reusable research components. I care about understanding which components contribute, and accounting for stability and compute alongside reward.
Make the idea concrete
Take an idea from equations to a working implementation. Separate models, data pipelines, training, and evaluation.
Build for reproducibility
Control randomness, track experimental settings, and design TensorFlow code with graph execution in mind.
Question the result
Verify baselines before adding complexity. Use ablations to investigate what actually contributes.
ON MY WORKBENCH
Research meets engineering.
Tools and setup from my profile.
Machine learning & research
Programming & systems
MLOps & data workflows
Hardware & creative tools
MY SETUP · AS DOCUMENTED
Training agents. Debugging kernels.
And occasionally remembering the GPUs can run games too.
- OS
- Manjaro Linux
- GPU
- 2 × RTX 4060 Ti
- Memory
- 128 GB RAM
- Storage
- NVMe SSD
OPEN TO RESEARCH DISCUSSIONS & COLLABORATIONS
Working on a
similar question?
If you’re thinking about exploration, diffusion models, or learning agents, I’d love to exchange ideas.
shuvrajeet17@gmail.com ↗