cv

Machine Learning Engineer

Kloudspot

Enhancing LISA AI by incorporating multi‑modal generative AI capabilities and delivering end‑to‑end production features.

• Enhanced LISA AI with advanced generative AI features and built an MCP client for seamless integration with enterprise data sources and tools.
• Deployed and optimized multiple LLMs on RTX-4090 GPUs, improving scalability to handle high-concurrency workloads with lower latency.
• Instruction finetuned multiple Open-source LLMs such as Llama 3.1, Deepseek, Qwen for advanced reasoning tasks and instruction alignments.
• Performed in-depth tuning of vLLM and SGLang frameworks with EAGLE/EAGLE-3 speculative decoding, achievingup to 2X higher throughput and 2.7–3.5X latency reduction in production.
• Tech stack: - Python, Docker, MongoDB, FastAPI, MCP, vLLM, SGLang, GPU Optimizations, PyTorch, Supervised fine-tuning (SFT), Large Language Models.

Associate Data Scientist

Celebal Technologies

Designed and deployed AI solutions across RAG, computer vision and predictive analytics on Azure with strong security and observability.

• Led AI solutions using Azure OpenAI and document intelligence to extract, embed and index structured/unstructured data; +15% BLEU and +10% semantic similarity.
• Implemented enterprise security with Azure AD, APIM and versioning across AI solutions.
• Improved RAG for tabular data; search accuracy 60%→85%; integrated chat history and logging via Azure Cosmos DB.
• Built hybrid VLM inference: quantized finetuned model on edge + full finetuned model on cloud; real‑time video analytics pipeline with adaptive edge–cloud switching; achieved 77.9% mAP (YOLO11, QLoRA, PyTorch, SmiolVLM).
• Developed predictive analytics for energy demand forecasting using ARIMA/SARIMA, XGBoost, LightGBM, CatBoost, K‑Means; achieved 75.2% R² for electricity load detection.
• Delivered on‑prem OCR using open‑source models and domain finetuning: Microsoft Table Transformer, Tesseract OCR, Layout Parser, Detectron2, ONNX Runtime.

Data Science Intern

Celebal Technologies

Improved enterprise search pipelines and finetuned LLMs for SQL/NoSQL tasks.

• Enhanced Azure OpenAI RAG for tabular data; accuracy 60%→85%; integrated chat/logging using Azure Cosmos DB.
• Finetuned GPT‑3.5‑turbo‑1106 and Mistral‑7B for SQL, NoSQL and Mongo query generation; >95% DataComp score.

Graduate Research Student

IISER Kolkata — Environmental Nanoscience Lab

Quantified estuarine metals and analyzed surface‑groundwater interactions with risk indices and correlation analysis.

• Quantified As/Mn in estuarine compartments; exceeded WHO limits (As 0.01 mg/L, Mn 0.05 mg/L).
• Evaluated soil/sediment toxicity and surface‑groundwater interactions.

Master of Science (MS)

Indian Institute of Science Education and Research Kolkata

Earth Sciences (Minor in Physics)

Bachelor of Science (BS)

Indian Institute of Science Education and Research Kolkata

Science (Physics, Mathematics, Earth Sciences)

U‑Net Paper Implementation (Medical Imaging)

Implemented U‑Net in PyTorch to segment 512×512 CXR images from a Kaggle pneumonia dataset.

• Dice score 0.87 after 3 epochs with a 23‑layer encoder‑decoder network.
• Masked images; batch size 16; Adam lr=0.0001; BCEWithLogitsLoss; local visualization.

Deep Convolutional GAN (DCGAN) on MNIST

Built a DCGAN in PyTorch to generate 28×28 MNIST digits from 60k images.

• FID 32.4 after 5 epochs using strided/transposed convolutions.
• Batch size 128; Adam lr=0.0002; BCE loss; batch norm; ReLU/LeakyReLU; TensorBoard visualization.