$whoami

APPLIED SCIENTIST & GOOGLE DEVELOPER EXPERT (GenAI)

High-impact Applied Scientist and Machine Learning Engineer with 5+ years of experience researching and architecting distributed LLM pre-training frameworks, 4D/5D-parallel training pipelines, and low-latency GenAI inference infrastructure. Recognized Google Developer Expert (GDE) in Generative AI and Hugging Face Fellow with a proven track record of scaling foundation models across multi-node GPU and Cloud TPU clusters, accelerating inference throughput by 40%+, and reducing cloud compute costs by 50%. Deep expertise in PyTorch, JAX/XLA, sharded parallelism (DP, FSDP/ZeRO-3, TP, SP, PP, CP), linear-time $O(N)$ sequence architectures, and production MLOps.

TECHNICAL SKILLS

• Distributed Training & Parallelism: 4D/5D Parallelism, Data Parallelism (DP), Fully Sharded Data Parallel (FSDP / ZeRO-3), Tensor Parallelism (TP), Sequence Parallelism (SP), Pipeline Parallelism (PP / GPipe / 1F1B), Context Parallelism (CP / Ring Attention), XLA / GSPMD, NCCL Collectives (all-reduce, all-gather, reduce-scatter, collective-permute).
• Deep Learning & Frameworks: PyTorch, JAX, Equinox, Optax, Orbax, TensorFlow, Keras, Hugging Face Transformers, Scikit-Learn, XGBoost, LightGBM.
• LLM Systems & Architecture: Pre-training & Post-training, RoPE, SwiGLU, Grouped-Query Attention (GQA), QK-Norm, KV-Cached Autoregressive Decoding, Activation Rematerialization (Checkpointing), Mixed Precision (bf16/fp16), FlashAttention, Fine-tuning (PEFT/LoRA/QLoRA), RAG Pipelines, vLLM, LlamaIndex, LangChain.
• Languages & Accelerators: Python, Mojo, CUDA, Pallas/Triton, Multi-GPU Nodes, Cloud TPU Pods, TensorRT.
• MLOps, Cloud & Databases: Google Cloud Platform (GCP), AWS, Docker, Kubernetes, FastAPI, Weights & Biases (W&B), MLflow, Perfetto Profiler, Redis, PostgreSQL, MongoDB, CI/CD.

PROFESSIONAL EXPERIENCE

Zomato (Eternal) | Applied Scientist II (Research) | Sep 2026 – Present

• Spearhead post-training and alignment for dark.ai, fine-tuning massive foundation models on a single 8x NVIDIA B200 (Blackwell) node by orchestrating FSDP/ZeRO-3, Tensor/Expert Parallelism (TP/EP), and Sequence Parallelism (SP) with FP8/BF16 mixed precision to fit full training states within 1.44TB HBM3e.
• Engineered scalable SFT and preference-optimization (RLHF/DPO) pipelines to steer conversational readability, tone, and targeted sycophancy/persona calibration for dark.ai, significantly boosting multi-turn user engagement and human-preference win rates.
• Maximized Model FLOPs Utilization (MFU) and reduced per-device HBM footprints by up to 75% via selective activation rematerialization and overlapping Blackwell Tensor Core compute with asynchronous NVLink 5 / NCCL collectives (all-gather, reduce-scatter, all-to-all), alongside high-throughput FP8 KV-cached inference serving.

IntelliTek Products Pvt. Ltd. | Software Engineer (Machine Learning) | Sep 2024 – Aug 2026

• Led Generative AI initiatives for clinical healthcare systems, automating structured clinical documentation (SOAP notes) and multi-turn diagnostic extraction from patient-doctor audio streams.
• Engineered HIPAA-compliant guardrails and PII/PHI redaction pipelines achieving 99.9% recall on clinical evaluation benchmarks.
• Optimized local LLM serving via vLLM, cutting external cloud API spend by 32% while maintaining sub-200ms Time-To-First-Token (TTFT) for streaming SOAP note generation.

TensorLake Inc. | Machine Learning Engineer | Apr 2024 – Aug 2024

• Architected Indexify, an open-source, real-time structured extraction engine designed for high-throughput multimodal unstructured data (PDFs, Audio, Video, and Images).
• Implemented hybrid dense-sparse retrieval and cross-encoder re-ranking pipelines that reduced production RAG query latency by 25%.
• Boosted distributed data ingestion throughput by 30% by refactoring legacy Python workflows into asynchronous, fault-tolerant microservices.

Dynopii Inc. | Machine Learning Engineer | Apr 2021 – Mar 2024

• Built end-to-end low-latency conversational AI and speech-to-intent pipelines, driving a 35% increase in automated customer resolution rates.
• Reduced distributed model training costs by 50% by engineering automated hyperparameter tuning workflows and dynamic AWS Spot-cluster orchestration.

PrediQt Business Solutions | Senior AI/ML Engineer (Contract) | Jun 2023 – Nov 2023 (Concurrent)

• Directed enterprise LLM fine-tuning and deployment initiatives, accelerating inference throughput by 40% via Post-Training Quantization (PTQ) and TensorRT execution graph optimization.
• Designed horizontally scalable model-serving architectures capable of sustaining 50% higher concurrent request volume during peak traffic spikes.

TensorFlow (Google) | Google Summer of Code Contributor | May 2022 – Sep 2022 (Concurrent)

• Authored official production benchmarks and architectural guides for TensorFlow Decision Forests (YDF), enabling scalable training, interpretation, and serving of Random Forests and GBDTs.
• Drove library adoption across the global Kaggle and competitive data science ecosystem by accumulating 15,000+ notebook views and 1,200+ forks.

Celebal Technologies | Data Scientist Intern | Sep 2021 – Dec 2021 (Concurrent)

• Engineered classical machine learning and NLP feature-extraction pipelines, accelerating algorithmic execution efficiency by 25% through vectorized Python and NumPy operations.

KEY PROJECTS & OPEN SOURCE

Jaxotron (4D/5D-Parallel LLM Pre-training & Inference Engine in JAX) | Creator

• Engineered a minimal, hackable 4D/5D-parallel (DP × FSDP/ZeRO-3 × PP × CP × TP + SP) LLM pre-training and inference framework in pure JAX and XLA GSPMD (~2,500 LOC; <400 LOC parallelism core) supporting Llama 3.x, Qwen3, Qwen2.5, and SmolLM2.
• Verified exact numerical parity (< 1e-5 max gradient delta and identical logits) across 8 combinable distributed sharding layouts on NVIDIA GPUs, Cloud TPUs, and simulated multi-device CPU meshes, reducing per-device memory by 4x+ via zero-copy sharded initialization.
• Built topology-independent asynchronous multi-host checkpointing (Orbax), direct Hugging Face .safetensors streaming with zero PyTorch dependency, XLA static collective inspection, and KV-cached tensor-parallel autoregressive decoding.

FireRequests (High-Performance Async Python Library) | Creator

• Architected an asynchronous file transfer and HTTP client library using Python asyncio and aiohttp, achieving 10x faster throughput for large model weights and dataset downloads compared to standard synchronous clients.
• Adopted in production environments by organizations including Roboflow for high-throughput multimodal dataset ingestion.

AutoTrain Advanced (Hugging Face) | Contributor

• Integrated Weights & Biases (W&B) experiment tracking, system telemetry, and model artifact versioning into Hugging Face’s official AutoTrain framework.
• Streamlined distributed LLM fine-tuning observability for thousands of ML practitioners across the open-source community.

AutoLGBM (Automated Tabular ML Library) | Creator

• Developed an automated gradient boosting framework combining LightGBM with Optuna Bayesian hyperparameter optimization, reducing validation log-loss by 12–18% over default baselines.
• Containerized inference endpoints using FastAPI and Docker, reducing production deployment overhead from hours to under 10 minutes.

PUBLICATIONS

• Context-Aware Music Retrieval with Yambda and JAX Data Parallelism – Official Keras Tutorial (KerasRS) | April 2026
• Gated Associative Memory: A Parallel O(N) Architecture for Efficient Sequence Modeling – arXiv:2509.00605 [cs.CL] | September 2025

EDUCATION

Doctor of Business Administration (DBA) | ESGCI, Paris | Expected 2028

• Research Focus: Business Model Innovation for Large Language Model Adoption in Clinical Workflows.

Bachelor of Technology (B.Tech) in Computer Science & Engineering | Netaji Subhash Engineering College, Kolkata | Graduated 2022

• CGPA: 8.9/10

AWARDS & COMMUNITY LEADERSHIP

• Google Developer Expert (GDE): Recognized by Google for technical leadership and expertise in Generative AI and Machine Learning.
• Hugging Face Fellow: Selected for impactful code and educational contributions to the global open-source NLP/LLM ecosystem.
• Community Co-Organizer: TensorFlow User Group (TFUG) & GDG Cloud Kolkata.