Cutdown
- Multimodal video-clip ranker combining transcript scoring, frozen embeddings, and a trainable ranker head
- Full FastAPI + Next.js upload-to-export pipeline evaluated against human-rated highlight clips
FastAPI · Next.js · FFmpeg · PyTorch
Selected projects
04 / 11A few things I've built, from multimodal video to the kernels underneath.
Forge
- Fused GPT-2 attention kernels in Triton, delivering 1.83x forward-pass speedup and 2x memory bandwidth
- Profiled with NVIDIA NCU to improve pipelining, prefetch stages, and SRAM reuse
Triton · CUDA · PyTorch · A100
InferenceBench
- Benchmarking harness comparing vLLM, SGLang, TensorRT-LLM under matched conditions
- Normalizes latency, throughput, and memory across five backends into an interactive dashboard
Python · vLLM · SGLang · TensorRT-LLM
Aegis
- Incident-remediation platform ingesting signals from 4 sources via hybrid pgvector + OpenAI embeddings
- Human-gated remediation workflows via Temporal, executing scoped Kubernetes actions with full audit trails
Next.js · Fastify · pgvector · Temporal · OpenAI · Kubernetes
Seven more experiments
Shadow Eval
- Production shadow-evaluation service replaying live requests against candidates without affecting users
- Blinded pairwise evaluation, statistical regression gates, PII redaction, and cost tracking; flagged an injected regression across 80 paired evaluations at p = 8.27×10⁻²⁵
Python · FastAPI · SQLite · Docker
LiveCanvas
- Self-hosted real-time collaborative whiteboard, drawing/cursors/presence synced via Yjs CRDTs
- Offline edits persisted to IndexedDB, auto-merging on reconnect across Redis-backed replicas
TypeScript · React · Yjs · Redis · PostgreSQL · Docker
PrivateText
- Fine-tuned PII-redaction classifier compiled with Torch-TensorRT FP16, cutting P50 latency 4.7x at 90.1% F1
- Deployed on Modal, serving character-accurate entity spans in production
PyTorch · Torch-TensorRT · FastAPI · Modal
Triage
- Fine-tuned ticket-priority classifier, quantized to INT8 ONNX, cutting latency 32.75ms → 17.15ms
- FastAPI microservice on Modal, 53.3% macro F1 on 6,000+ class-imbalanced tickets
PyTorch · ONNX Runtime · FastAPI · Modal
AskMyDB
- Text-to-SQL copilot using OpenAI function calling with read-only SELECT validation, 93% accuracy
- Dockerized FastAPI + React + Postgres, auto-rendering charts, deployed on Railway
FastAPI · OpenAI · PostgreSQL · React
Surge
- Extended vLLM's paged-attention serving for Llama-3-8B, increasing throughput 2.4x at 100ms p99 latency
- Optimized in-flight batching and KV-cache eviction under bursty single-GPU load
vLLM · CUDA · Kubernetes
Squelch
- Post-training INT8 quantization on Whisper-small via TensorRT, cutting latency 2.7x at 0.3 WER degradation
- Profiled memory bandwidth and SM occupancy with NVIDIA Nsight to remove throughput regressions
PyTorch · TensorRT · ONNX
Experience
06 entriesResearch and applied work, from undergraduate study to hands-on robotics and ML systems.
Software Engineer
Eminent Services Corporation
- Tested IDMS product workflows and backend integrations, identifying functional defects across production features
- Collaborated with developers to reproduce failures, document findings, and support root-cause analysis and fixes
QA · Backend
ML & Robotics Researcher
NYU CILVR (GRAIL)
- Developed tactile-centric world models using AnySkin sensor + video for reorientation tasks in LIBERO
- Designed a custom "push-to-align" block–cabinet task with full dynamics and latent-space modeling for RL
- Built a tactile/proprioceptive data pipeline; fixed a temporal aggregation bug, improving held-out completion
- Deployed robotics workloads via Singularity with Tailscale-based remote access, enabling off-site experiment execution
PyTorch · MuJoCo · Robosuite · RL · Singularity
ML Systems & Inference Researcher
System & AI Lab (NYU)
- Reduced Efficient-VAR inference latency 17% on NVIDIA A100s via Triton kernels and memory-path profiling
- Built a non-invasive instrumentation framework to visualize top-k / top-p / CFG tradeoffs
- Shipped transfer-learning pipelines across 3 research projects; PostgreSQL experiment tracking cut logging overhead 35%
- Orchestrated multi-node GPU experiments on Slurm, debugging training failures and scheduling overnight evaluations
Triton · CUDA · PostgreSQL · Slurm
Machine Learning Intern
Tata Steel
- Built a generative AI system for automated steel surface defect detection
- GAN-based data generation pipeline lifted rare-class recall by 28%; containerized training-to-inference with Docker + GitHub Actions
GANs · Docker · CI/CD
Research Intern
Samsung R&D Institute India
- Optimized CNN/RNN architectures for <10ms mobile inference via quantization and pruning
- Built production pipelines syncing vision, audio, and sensor signals at sub-100ms end-to-end latency
Quantization · Pruning · Mobile
Data Science Intern
Hewlett Packard Enterprise
- Trained multilingual text classifiers on 11k+ samples (93% accuracy) with custom NLP pipelines
- Deployed distributed training on Azure ML with hyperparameter tuning and A/B testing infrastructure
NLP · Azure ML
Open source
03 entriesSmall changes in systems I care about. The pull requests tell the story.
- Added input validation and regression tests for distributed checkpointing and schedulers, preventing runtime failures (PR #8126)
- Built optimization-pass contracts and schedule validation for DeepCompile, detecting pre-execution conflicts (PR #8139)
Python · PyTorch · Distributed Systems
- Root-caused and patched an out-of-bounds memory read in a Triton MoE kernel, resolving a CUDA illegal access (PR #48371)
- Developed a regression test suite, resolving 12/60 failing cases with up to 98% reference-implementation mismatch
Python · Triton · CUDA
- Added a task-prompt configuration option to RoboCasa's eval env, fixing a natural-language/CamelCase mismatch (PR #4073)
- Developed unit tests covering task-prompt selection and configuration validation
Python · PyTorch · Gym
Publications & thesis
05 entriesReference Wording Shapes Retrieval of Robot Task Instructions
arXiv preprint · arXiv:2610.17549
Studies how reference wording changes instruction-to-task retrieval in RLBench. Across 62 tasks and 80 reference galleries, character TF-IDF reaches 61.3% mean top-1 accuracy; adding a second reference description improves accuracy by 7.0 percentage points on matched queries.
Visuo-Tactile World Models for Dexterous Manipulation: From Simulation Benchmarks to Real-World Skin Sensing
MASTER'S THESIS · NEW YORK UNIVERSITY · 2026
Builds world models that fuse tactile (AnySkin) and visual signals for dexterous robotic manipulation, evaluated across simulation benchmarks and real-world skin-sensing setups.
2025
Prediction of Cerebrospinal Fluid (CSF) Pressure with Generative Adversarial Network Synthetic Plasma-CSF Biomarker Pairing
FIG. 01 · NEUROINFORMATICS · VOL 23(3) · 2025
Introduces a GAN-augmented ensemble model using plasma cfRNA to predict CSF pressure non-invasively. Demonstrates high predictive accuracy and clinical relevance for ICP monitoring in spaceflight and clinical contexts.
2025
Performance Evaluation of Generative Adversarial Networks for Anime Face Synthesis Using Deep Learning Approaches
FIG. 02 · MULTIDISCIPLINARY SCIENCE JOURNAL · VOL 7(3) · 2025
Comparative study of GAN variants (DCGAN, ProGAN, StyleGAN2) for anime face synthesis. StyleGAN2 demonstrated superior effectiveness, achieving the lowest FID score among tested models.
2024
Artificial Intelligence-Based Methodologies for Early Diagnostic Precision and Personalized Therapeutic Strategies in Neuro-Ophthalmic and Neurodegenerative Pathologies
FIG. 03 · BRAIN SCIENCES · VOL 14(12) · 2024
Review highlighting the role of AI-enhanced diffusion MRI and PET imaging in early diagnosis of neurodegenerative and neuro-ophthalmic disorders, integrating multimodal data with deep learning models for improved precision.
A little more context
2020 — 2026
Between models
and machines.
I'm drawn to the parts of ML that meet the real world: a robot learning through touch, a model waiting on memory, an experiment that needs to run reliably.
Six years of computer science and computer engineering have taken me through research, applied ML, and the systems underneath. I like being able to follow the problem across those boundaries.
Right now, I'm digging further into how hardware shapes a model, how to understand what it learns, and how to keep agents useful over longer tasks.
New York University
Master's in Computer Engineering · GPA 3.83/4.0
Coursework: ML, Computer Vision, Reinforcement Learning, Hardware Accelerator Design. Thesis: Visuo-Tactile World Models for Dexterous Manipulation.
Vellore Institute of Technology
Bachelor's in Computer Science Engineering · GPA 3.97/4.0
Visiting student at the National University of Singapore, Dec 2022 – May 2023. Also taught as Student Instructor for CSE1006 (Blockchain & Cryptocurrency Technologies, Winter 2023) and CSE4001 (Parallel & Distributed Computing, Fall 2023).
Tools I work with
Inference & Optimization
Triton · TensorRT · vLLM · KV-cache · Paged Attention
Languages
Python · C/C++ · CUDA · SQL · TypeScript · Bash · MATLAB
ML & DL
PyTorch · JAX · TensorFlow · scikit-learn · OpenCV · RAG · LangChain · LangGraph
Infrastructure & Backend
FastAPI · Node.js · AWS · GCP · Azure ML · Docker · Kubernetes · CI/CD · PostgreSQL · pgvector · MongoDB · Pinecone
Tools
Git · Linux · Claude Code · Codex
Robotics & RL
ROS · MuJoCo · Robosuite · IsaacSim · Stable-Baselines3
Good problems. Good people.
I'm interested in ML engineering and research roles, thoughtful teams, and ideas worth testing.
Let's talk ↗