Kartik Kuppam profile photo 1 on Tal

Kartik Kuppam

Lead Data Scientist | Samsung R&D | Ex-Bosch | 9+years | Multi-Model | Computer Vision | NLP/Gen AI | DL | ML

Bengaluru

10+ years of work experience

Samsung R&D Institute India - Bangalore
Aug 20259 mosPresent

Bengaluru, Karnataka, India

Bosch Research
Jul 20232 yrs 1 moJul 2025

LLM-Driven Image Captioning & Retrieval for Autonomous Driving Deployed a LoRA-tuned CLIP system for text-to-image retrieval in plant environments, cutting inspection time and driving €1M+ in productivity gains. Constructed a domain-specific dataset by pairing segmentation metadata with contextual captions, incorporating 256-token expansion, hard negative mining, and angular margin loss for contrastive learning. Replaced GPT-based generation with fine-tuned LLaMA-1B/7B models using LoRA and QLoRA, boosting ROUGE-L by 15% and reducing GenAI inference costs by 80%+. AskBosch: Agentic RAG Chatbot with GPT-4o Designed and deployed an agentic Retrieval-Augmented Generation (RAG) system to process 2TB+ of Bosch internal documents. Built a scalable backend with Azure Functions + Databricks and an interactive Streamlit front-end, enabling internal rollout via the AskBosch GenAI suite—projected to save €4M annually for 1,000+ users. Enhanced document ranking using multi-query expansion, dense+BM25 hybrid retrieval, Reciprocal Rank Fusion, and bi-encoder re-ranking. Used LLaMA 3.2-11B on Databricks for structured insight extraction.
Bosch Research
Jan 20212 yrs 7 mosJul 2023

AOI Studio – Foundation Models for Manufacturing Developed self-supervised foundation models using ViT and ResNet-50, leveraging SimCLR and DINO for feature learning on unlabeled manufacturing data. Achieved 90% accuracy with just 20–30% labeled data—outperforming transfer learning baselines with 60–70% less supervision. Enabled €1M in savings through reduced annotation effort and faster deployment. Integrated into AOI Studio, reducing labeling and training costs by 30–40%, and scaling across multiple Bosch manufacturing use cases. Applied knowledge distillation and diffusion-based generative models (DDPM, DDIM) to improve visual inspection robustness. Regulation Document Classification Built a scalable NLP pipeline for long document classification using BERT and DistilBERT, automating compliance triage for over 60K regulatory docs across 4 Bosch units. Engineered metadata-based features and optimized for 99% recall and 63% precision, addressing class imbalance and label ambiguity. Delivered €10M+ annual cost avoidance via automation.
Bosch Center for Artificial Intelligence (BCAI)
Jul 20191 yr 7 mosJan 2021

Bengaluru, Karnataka, India

Developed and deployed AI-powered defect detection and segmentation models (YOLOv3, YOLOv4, U-Net) with 150ms latency, optimizing accuracy and real-time performance in manufacturing. Optimized and converted models to ONNX, deploying via TorchServe on WSL, reducing inspection costs by €300K annually. Designed a radar data generation pipeline, simulating real-world sensor data using PointNet, PointPillars, and PointCapsuleNet. Trained Autoencoders (AE) and Variational Autoencoders (VAE) for point cloud data synthesis, improving anomaly detection using MSE and Mahalanobis Distance.
DRDO, Bangalore
Jun 20172 yrs 1 moJun 2019

Bengaluru, Karnataka, India

Designed a real-time system using Conv3D, MobileNet, and RNN, trained on custom video datasets for display control in tactical applications. Developed an NLP-driven pipeline for classifying technical documents and enabling SQL-based audit trails, improving traceability and cutting manual review time by 60%.
Indian Oil Corporation Limited
Oct 20151 yr 8 mosMay 2017

Bhubaneshwar, Odisha, India

GitHub
@kuppamk

Kartik’s AI Stack

gemini
gemini
langchain
langchain
langgraph
langgraph
openai
openai

Education

Liverpool John Moores University

Master's degree, Artificial Intelligence

Liverpool John Moores University

20191 yr2020
International Institute of Information Technology Bangalore

PG Diploma, Artificial Intelligence

International Institute of Information Technology Bangalore

20181 yr2019

Join Kartik and 10k+ other professionals on

Tal Boss

Tal is the app where bosses meet engineers & builders directly, without middlemen

Download Tal Boss

Kartik Kuppam on Tal

Headline

Lead Data Scientist | Samsung R&D | Ex-Bosch | 9+years | Multi-Model | Computer Vision | NLP/Gen AI | DL | ML

About

As a Lead Data Scientist at the Bosch Research, my role revolves around pioneering AI-driven solutions in manufacturing and engineering. Our team has successfully delivered projects that reduce labeling and training costs by up to 40%, thanks to innovative foundation models and self-supervised learning techniques. In collaboration, we have fine-tuned large language models for enhanced image retrieval, achieving a ROUGE-L score of 95% and realizing annual savings of €1M. Our latest triumph involved optimizing defect classification using state-of-the-art models like ResNet-50 and CLIP, combined with knowledge distillation, to tackle class imbalance and cut inspection costs by €300K annually. With my expertise in generative AI, convolutional neural networks, and deployment tools like ONNX and TorchServe, I am committed to pushing the boundaries of AI application and contributing to impactful, cost-effective solutions in the industry.

Location

Bengaluru

AI stack

gemini

langchain

langgraph

openai

Work experience

Lead Data Scientist at Samsung R&D Institute India - Bangalore, Aug 2025 - Present · 9 mos, Bengaluru, Karnataka, India

Lead Data Scientist at Bosch Research, Jul 2023 - Jul 2025 · 2 yrs 1 mo, LLM-Driven Image Captioning & Retrieval for Autonomous Driving Deployed a LoRA-tuned CLIP system for text-to-image retrieval in plant environments, cutting inspection time and driving €1M+ in productivity gains. Constructed a domain-specific dataset by pairing segmentation metadata with contextual captions, incorporating 256-token expansion, hard negative mining, and angular margin loss for contrastive learning. Replaced GPT-based generation with fine-tuned LLaMA-1B/7B models using LoRA and QLoRA, boosting ROUGE-L by 15% and reducing GenAI inference costs by 80%+. AskBosch: Agentic RAG Chatbot with GPT-4o Designed and deployed an agentic Retrieval-Augmented Generation (RAG) system to process 2TB+ of Bosch internal documents. Built a scalable backend with Azure Functions + Databricks and an interactive Streamlit front-end, enabling internal rollout via the AskBosch GenAI suite—projected to save €4M annually for 1,000+ users. Enhanced document ranking using multi-query expansion, dense+BM25 hybrid retrieval, Reciprocal Rank Fusion, and bi-encoder re-ranking. Used LLaMA 3.2-11B on Databricks for structured insight extraction.

Senior Data Scientist at Bosch Research, Jan 2021 - Jul 2023 · 2 yrs 7 mos, AOI Studio – Foundation Models for Manufacturing Developed self-supervised foundation models using ViT and ResNet-50, leveraging SimCLR and DINO for feature learning on unlabeled manufacturing data. Achieved 90% accuracy with just 20–30% labeled data—outperforming transfer learning baselines with 60–70% less supervision. Enabled €1M in savings through reduced annotation effort and faster deployment. Integrated into AOI Studio, reducing labeling and training costs by 30–40%, and scaling across multiple Bosch manufacturing use cases. Applied knowledge distillation and diffusion-based generative models (DDPM, DDIM) to improve visual inspection robustness. Regulation Document Classification Built a scalable NLP pipeline for long document classification using BERT and DistilBERT, automating compliance triage for over 60K regulatory docs across 4 Bosch units. Engineered metadata-based features and optimized for 99% recall and 63% precision, addressing class imbalance and label ambiguity. Delivered €10M+ annual cost avoidance via automation.

Data Scientist at Bosch Center for Artificial Intelligence (BCAI), Jul 2019 - Jan 2021 · 1 yr 7 mos, Bengaluru, Karnataka, India, Developed and deployed AI-powered defect detection and segmentation models (YOLOv3, YOLOv4, U-Net) with 150ms latency, optimizing accuracy and real-time performance in manufacturing. Optimized and converted models to ONNX, deploying via TorchServe on WSL, reducing inspection costs by €300K annually. Designed a radar data generation pipeline, simulating real-world sensor data using PointNet, PointPillars, and PointCapsuleNet. Trained Autoencoders (AE) and Variational Autoencoders (VAE) for point cloud data synthesis, improving anomaly detection using MSE and Mahalanobis Distance.

Scientist-B at DRDO, Bangalore, Jun 2017 - Jun 2019 · 2 yrs 1 mo, Bengaluru, Karnataka, India, Designed a real-time system using Conv3D, MobileNet, and RNN, trained on custom video datasets for display control in tactical applications. Developed an NLP-driven pipeline for classifying technical documents and enabling SQL-based audit trails, improving traceability and cutting manual review time by 60%.

Project Engineer at Indian Oil Corporation Limited, Oct 2015 - May 2017 · 1 yr 8 mos, Bhubaneshwar, Odisha, India

Education

Master's degree, Artificial Intelligence, Liverpool John Moores University

PG Diploma, Artificial Intelligence, International Institute of Information Technology Bangalore

Public links

github: https://github.com/kuppamk