About
Arizona State University
VIT Bhopal University
Robot Learning & Machine Learning Systems Researcher
Computer Science master's student working on assembly planning for dual-arm robots, and on making large models train and run faster.
I'm finishing my Master's in Computer Science at Arizona State University, where my thesis at RISE Labs (advised by Dr. Wenlong Zhang) plans robotic assembly sequences for a 6-piece burr puzzle on a dual-arm ABB YuMi — a problem that is non-monotone and has zero clearance at insertion, so it breaks the assumptions most pick-and-place planners are built on. In parallel, at ASU's Knowledge Discovery & Data Mining group, I work on diversity and mode collapse in LLM math reasoning and on simulation-to-decision frameworks, and I co-authored DELTA (ICDM 2025) on privacy-preserving data reprogramming.
The thread running through it is that I like problems where the geometry, the learning and the systems all have to be right at once: writing a Triton FlashAttention kernel and measuring it against the memory roofline, auditing a robot teleoperation dataset frame by frame before fine-tuning a 7B vision-language-action model on it, or getting a planned sequence off the screen and onto a real robot. Before graduate school I spent two years at Shell as an Analytics Engineer, where a computer-vision lube-monitoring system I worked on is patent-pending and serves 500+ ships daily.
Technical Expertise
Research Areas
Languages
Robotics & Simulation
ML Systems & Cloud
Experience
Research Experience
Graduate Research Assistant — Thesis
RISE Labs, Arizona State University
January 2025 - Present
Tempe, AZ, USA · Advised by Dr. Wenlong Zhang
- Building an assembly planning pipeline for multi-stage interlocking assembly, using a 6-piece burr puzzle as a proxy for parts that are non-monotone to assemble and have zero clearance at insertion.
- Introduced a vertex-based action space that defines moves as corner-to-corner alignments, restricted by swept-volume collision checks and bottom-support constraints that model one-gripper, two-gripper and unlimited-support budgets.
- Designed a value approximation heuristic with pruned n-step lookahead and greedy rollout, applied assembly-by-disassembly to cut the search space, and validated the resulting pick-and-place sequence in ABB RobotStudio with RAPID motion programming before transferring it to the physical ABB IRB 14000 YuMi.
- Presented initial results as Dynamic Programming in Cooperative Autonomous Assembly (with Max Gao) at the Southwest Robotics Symposium, November 2025. Now extending the framework toward full sim-to-real bimanual manipulation with robot dynamics and camera-based perception.
Graduate Research Assistant
Knowledge Discovery & Data Mining Group, Arizona State University
March 2025 - Present
Tempe, AZ, USA
- Investigating mode collapse in LLM math reasoning: built a synthetic-data generation pipeline with vLLM and a GRPO training environment (Qwen3-4B), analysing output diversity, reasoning trajectories and error patterns to improve dataset quality and model robustness.
- Developed a simulation-to-decision framework for supply chain optimization, integrating a Mixture Density Network, an LSTM-based simulator, and an LLM decision maker.
- Co-authored DELTA, a two-phase variational disentangled learning framework for privacy-preserving data reprogramming that combines reinforcement learning with a disentangled VAE — accepted at ICDM 2025.
Industry Experience
Associate Analytics Engineer
Shell Private Limited Company
Aug 2022 - Aug 2024
Bangalore, Karnataka, India
- Worked on large complex financial data using Alteryx and SQL databases to automate workflows which resulted in 40% reduction in the workload of the team. Built an interactive dashboard to provide a holistic view of the finances.
- Collaborated with a global team on a deep learning-based image analysis project for lube monitoring. It involved developing a python-based model for ROI detection and classification. Worked with Linux clusters to enhance workflow efficiency.
- Automated geo-mechanical experiments with python which lead to a drastic reduction in the time to generate reports by 50% per lab technician.
- Developed a process automation framework for business-critical deployments for disaster recovery involving more than 1000 applications. This led to a reduction in the workload of 3 business days per team member.
- Developed an interactive dashboard to track 150+ business continuity services, including disaster recovery and enterprise recovery for business-critical applications.
- Collaborated with 20+ application owners, stakeholders, and vendors to ensure the seamless execution of the disaster recovery process being a disaster recovery focal point.
Publications
My recent research publications in peer-reviewed conferences and journals.
DELTA: Variational Disentangled Learning for Privacy-Preserving Data Reprogramming
ICDM - International Conference on Data Mining, 2025
A two-phase variational learning framework for privacy-preserving feature transformation, combining reinforcement learning with a disentangled VAE. Demonstrated robust performance across eight real-world datasets, improving downstream task utility while reducing sensitive attribute leakage.
Presenting DELTA at ICDM 2025.
Advanced Deepfake Detection Using Inception-ResNet-v2
International Conference on Communication and Intelligent Systems, 2023
A deepfake detection pipeline combining MTCNN-based face extraction with Inception-ResNet-v2 and custom dense layers in a sequential model, reaching 91.41% accuracy on the DFDC dataset with an interactive GUI.
Patents
Intellectual property from my research and development work.
MITRA: Marine Inspection Tool for Rating and Assessment
Shell PLC | Filed: 2024
An innovative solution in lube monitoring that uses deep learning for image classification with 90% accuracy. The project - Shell LubeMonitor leverages deep-learning, high-resolution image analysis and a vast engine data archive to provide granular insights and early warnings. It is already helping over 500 ships daily stay ahead of performance issues.
Patent Pending
Projects
Robot Learning & Assembly Planning
Assembly Planning with the Burr Puzzle Thesis
January 2025 - Present · RISE Labs, ASU
A 6-piece burr puzzle is a hard assembly instance: it is non-monotone (pieces must be repositioned, not just placed once) and its clearances are zero at insertion. This thesis plans a full assembly sequence under both constraints for a dual-arm ABB YuMi, using a corner-to-corner action space and a pruned-lookahead search that plans the disassembly and reverses it. The sequence was validated in ABB RobotStudio and transferred to the physical robot, and presented at the Southwest Robotics Symposium 2025.
One-arm vs. two-arm assembly · the 3D-printed burr · presenting at SWRS 2025.
The action space
For each piece the planner extracts vertices where three mutually orthogonal edges meet, and an action translates the moving piece linearly until a chosen pair of corners coincides — a discrete, geometry-derived action space rather than a sampled one. Feasibility is decided by interpolating along the translation and collision-checking against every other active piece, plus a bottom-support test and an arm budget that models one gripper, two grippers, or unlimited support.
The search
From each state the branch tree is pruned at every level, expanded for an n-step lookahead, and each surviving branch finished with a truncated greedy rollout; the move with the best cumulative score is committed. Accepting short-term sub-optimal moves this way is what rescues the goal-distance heuristic from the interlocking structure that traps a purely greedy planner.
Assembly by disassembly
The sequence is searched backwards from the interlocked goal and then reversed, because disassembly is the easier direction: every removal moves away from a fully constrained pose, while assembly must find the one entry order that does not deadlock. Plans were generated for one-, two-, and unlimited-arm budgets, with more arms reducing the number of required actions.
From simulation to the robot
The pick-and-place sequence was validated in ABB RobotStudio with RAPID motion programming and transferred to the physical ABB IRB 14000 YuMi. A MuJoCo scene converted from the real robot's URDF (32 bodies, 60 DoF, 16 actuators) carries the six pieces as free bodies for headless rollout. Current work brings gripper access, arm collision and reach into the scoring itself, so a move no arm can perform is rejected at the stage that proposed it.
Southwest Robotics Symposium 2025
Presented Dynamic Programming in Cooperative Autonomous Assembly, independent research with Max Gao, arguing that manually sequenced or heuristic assembly plans do not scale to diverse assemblies and that interlocking connections need a formulation reasoning directly about geometry and support. A parallel study asks whether a language model, given solved sub-puzzle traces in context, can propose feasible moves without search — write-up in progress.
Robotic Chess Player using Foundation Models
May 2025 - August 2025
• Built an end-to-end synthetic data generation pipeline that converts chess games in PGN notation into robot demonstrations, using a state machine and operational space control with a UR5 manipulator to physically play the games on a board in MuJoCo.
• Fine-tuned the Pi0 vision-language-action foundation model on those synthetic demonstrations, and proposed an extension in which a single black-box model handles both game reasoning and low-level motion planning.
Door-Opening Policy with TD3
June 2025
• Trained a Twin Delayed DDPG (TD3) agent to open a door with a Franka Panda arm in the robosuite Door environment, under joint-velocity control at 20 Hz with a 300-step horizon.
• Implemented the agent from scratch in PyTorch — twin critics, delayed and soft target updates, target-action smoothing noise, and a fixed-size replay buffer — with checkpointing and TensorBoard logging over training runs of up to 10,000 episodes.
LLM Systems, Inference & Data
FlashAttention for Llama Inference, in Triton and CUDA
January 2025 - May 2025 · CSE 524, Machine Learning Acceleration
• Attention is memory-bound, not compute-bound: a naive implementation writes the full N×N score matrix out to HBM, reloads it for softmax, writes it back, and reloads it again for the second matmul — so the GPU idles on memory traffic while on-chip SRAM runs an order of magnitude faster than HBM.
• Implemented FlashAttention-style attention as a hand-written @triton.jit kernel that tiles Q/K/V into SRAM-sized blocks and fuses matmul → softmax → matmul into one kernel, using the online (streaming) softmax algorithm — a running max and normalizer per query row, with the output accumulator rescaled whenever the running max shifts — so the result is mathematically identical to standard softmax attention.
• Swapped that kernel into a Llama decoder block, keeping RMSNorm, RoPE, grouped multi-query attention, the KV cache, and SwiGLU untouched, and benchmarked against a naive baseline on a 15M-parameter Llama2-architecture model trained on TinyStories.
• Measured ~4× faster inference and >10× improvements in memory traffic, bandwidth utilization, and arithmetic intensity. Tested on ASU's Sol cluster (NVIDIA A100) and locally on an RTX 4060.
Latency, improvement factors, roofline placement and benchmark summary.
Perf-Analyze — Inference Performance Explorer Live demo
June 2025
• Built a tool that turns raw LLM inference performance sweeps (.xlsx projections from an AI-accelerator vendor) into decisions for two very different audiences, rather than a single table that serves neither.
• Customer / PM view — a go/no-go signal against user-defined requirements, a ranked throughput bar, a performance frontier scatter (throughput vs. time-to-first-token), and a model × workload profile heatmap.
• Internal engineer view — throughput distribution, relative metric comparison, scaling efficiency, config sensitivity curves, and a t-SNE cluster map over the sweep space.
• Eleven models pre-load at startup and any conforming sweep — including models the app has never seen — renders live on upload with no rebuild and no code changes. Built with Gradio, pandas and Plotly; deployed on Hugging Face Spaces.
Statistically Representative Benchmark Pruning
June 2025
• Reduced three evaluation benchmarks (LiveCodeBench, AA-LCR, MMMU) to small subsets that preserve model rankings, go/no-go decisions and the difficulty distribution of the full set.
• Stratified every item into easy/medium/hard by average pass rate, allocated the budget proportionally, then ran spectral clustering within each stratum over sentence-transformer embeddings of the problem text.
• Selected the centroid-nearest item in the easy and hard strata, and in the medium stratum the item with the highest IRT informativeness (a 2PL-inspired score, since a full fit is infeasible with three models) — because medium items are where model abilities separate most.
• Derived target sizes from a stratum-coverage floor and validated them against a sensitivity sweep on ranking preservation (Spearman ρ) and mean absolute error.
OpenArm 2.0 — Robot Data Pipeline & VLA Fine-Tuning
July 2025 - August 2025
• An end-to-end pipeline over lerobot/aloha_static_coffee — 50 bimanual teleoperated demonstrations, 55,000 frames at 50 fps, 14-D state/action/effort, four synchronised cameras (two third-person, two wrist-mounted).
• Audit. Five findings, of which only two justify deleting data — the discipline being to remove data only when the fix is unambiguous and bias-free, and otherwise annotate and defer. The headline measurement: video lags state by 5 frames (100 ms) in all 50 episodes.
• Labeling. A three-level schema (episode / segment / frame) with a weak labeler that proposes ~10–11 segments per episode with no human input, so annotators correct proposals instead of drawing boundaries from scratch; inter-annotator agreement measured at tIoU 0.589.
• Curation. A config-driven filter registry implementing the audit findings one-for-one, producing a valid LeRobotDataset at 79.7% frame retention, with 16 tests over alignment and synthetic corruption.
• Evaluation & fine-tuning. A three-tier evaluation protocol with a milestone success criterion (the dataset carries no success label, so the protocol has to create one), then OpenVLA-7B fine-tuned both with LoRA and full-parameter for a full epoch, split by episode — checkpoints published to the Hugging Face Hub with the action statistics needed to decode the model's bin indices back into joint targets.
One episode as a contact sheet · the measured video/state lag · retention through the curation filters · the segment annotator.
Mode Collapse in LLM Math Reasoning
March 2025 - Present · KDD Group, ASU
• Investigating how models trained on synthetic math data lose diversity: built a vLLM-based generation pipeline producing reasoning traces from a base model and from models fine-tuned separately on real and synthetic math data.
• Measured style collapse by skeletonizing traces (stripping numbers and variables to leave the reasoning form), embedding the skeletons, clustering them, and computing conditional entropy of style given the question, unique clusters per question, and top-1 cluster mass — so diversity is measured over how a model reasons, not what it answers.
• Built a GRPO reinforcement-learning training environment (Qwen3-4B) with a reward-logging harness to study how RL training shifts that diversity, alongside output and error-pattern analysis for improving dataset robustness.
Sim-to-Decision: MDN + LLM for Supply Chain Optimization
March 2025 - Present · KDD Group, ASU
• A simulation-to-decision framework that pairs a learned simulator with an LLM decision maker: a Mixture Density Network and an LSTM-based simulator model the stochastic outcome of an action, and the LLM proposes decisions evaluated against that simulated world.
• Modular training pipeline across four supply-chain datasets (including an out-of-distribution split), supporting simulator-only, decision-maker-only, or joint training, with zero-shot and sample-efficiency evaluations against RL baselines.
Computer Vision & Medical Imaging
Self-Supervised Dense Point Tracking in Turbulent Videos
August 2024 - December 2024
• Developed a benchmark for point tracking in videos with induced atmospheric turbulence, enhancing the DINO-Tracker with RAFT-based optical flow refinement.
• Diagnosed robustness gaps in DINO-Tracker, revealing a 32% average-Jaccard drop under moderate turbulence and mapping failure modes in occluded, high-blur frames.
• Estimated the strength of turbulence using the trajectories of points in the turbulent videos.
Swin Transformer for Vision Tasks
August 2024 - December 2024
• Used Swin Transformer for classification, localization, and segmentation on ChestXRay14, NODE21, and ChexMask datasets to optimize pipelines and improve performance in medical image analysis.
• Achieved 72.64% accuracy training a model from scratch and 81.78% accuracy fine-tuning on ImageNet for ChestXRay14 classification.
Comparative Study of Video Retrieval Methods
August 2024 - December 2024
• Implemented dimensionality reduction techniques (PCA, SVD, LDA, K-Means) on ResNet, HOG, HOF, and Color Histogram feature spaces to rank and retrieve videos based on similarity.
• Developed centroid-based ranking and label prediction models, achieving improved retrieval accuracy and demonstrating SVD's effectiveness in preserving latent structures.
Metamorphosis Automation
Aug 2020 - May 2021
• Developed a deep learning model using ResNet-34 and CNNs to detect car accidents from real-world CCTV surveillance footage with high accuracy.
• Built a complete pipeline integrating ML inference, a RESTful API, and Android/web dashboards to deliver real-time accident alerts with geo-tagged evidence.
Cloud & Edge Computing
Edge-Based Face Recognition Pipeline using AWS IoT Greengrass
April 2025 - May 2025
• Designed and deployed a distributed face recognition pipeline by integrating AWS IoT Greengrass, MQTT, Lambda, and SQS for low-latency edge processing.
• Implemented real-time face detection on simulated IoT edge devices using MTCNN and Python components deployed via custom Greengrass components.
• Enabled secure device-to-cloud communication by configuring AWS IoT Core policies, certificates, and MQTT bridges for message exchange.
• Triggered cloud-based FaceNet-based face recognition via SQS queues and Lambda functions, maintaining a privacy-preserving and scalable architecture.
Serverless Face Recognition Pipeline with AWS Lambda & ECR
March 2025 - April 2025
• Developed a face recognition service using AWS Lambda, Docker, and Elastic Container Registry (ECR) to enable serverless ML inference on streaming video frames.
• Built and containerized custom Lambda functions for face detection (MTCNN) and recognition (FaceNet) with PyTorch models, deployed via ECR.
• Integrated AWS SQS for decoupled communication between detection and recognition stages, ensuring scalability and low latency.
Cloud-Based Face Recognition System Architecture and Optimization
February 2025 - March 2025
• Designed a scalable cloud-based face recognition system on AWS using S3, SQS, and dynamic EC2 instances for efficient image processing.
• Built a responsive Python web tier to handle HTTP requests, manage cloud storage, and coordinate communication between the UI and processing layer.
• Engineered the application tier for deep learning inference with PyTorch, processing images from cloud storage and returning results via asynchronous message queues.
• Developed an auto-scaling controller to dynamically provision resources, optimizing costs and ensuring system responsiveness with zero idle instances.
Mobile & Web Development
mShare
August 2020
• Developed and published a file-sharing mobile application on the Google Play Store.
• Enabled high-speed offline sharing via hotspot and QR code with support for large files including photos videos, apps, and folders.
• Implemented advanced features such as pause/resume transfers, multi-device sharing, and automatic reconnection on network changes.
• Designed a lightweight, speed-optimized UI for reliable performance on low-end Android devices.
• Integrated network speed measurement and dynamic connection switching to enhance user experience.
•
GitHub Repository
COVFeed
Jun 2020
• Built a mobile application using the Flutter framework to serve as a semi-news platform for healthcare professionals.
• Designed for doctors, nurses, and medical staff to share real-time regional health updates during the COVID-19 pandemic.
• Developed in a team of two as part of the International Flutter Hackathon – Hack’20.
•
GitHub Repository
•
YouTube Video
HabitUp — Habit Tracking App
March 2022 - June 2022 · Software Development Engineer (Intern), Remote
• Built a habit tracking Android application with the Flutter framework and shipped it to the Google Play Store.
• The published app went on to support over 100,000 users tracking their habits.
ML Web Apps — Toxic Comments & Movie Ratings
February 2021 - April 2021 · Machine Learning Engineer (Intern), Technocolabs
• Toxic Comments Classification — a Flask web app serving a model that sorts comments into toxicity categories, deployed on Heroku.
• Movie Rating Prediction — a Flask web app predicting movie ratings, built as the major project while leading a team of five, also deployed on Heroku.
Contact
Feel free to reach out to me if you have any questions, opportunities, or just want to connect!
Location:
Tempe, AZ, USA