Headline

AI/ML - Lead @ kapture.cx | Voice AI, Generative AI

About

Graduated from the Indian Institute of Technology, Madras, with a Dual Degree in Engineering Design. Currently serving as AI/ML - Lead at Kapture CX. With expertise in Voice AI, Text-to-Speech, and Generative AI, they contribute to advancing the AI field as a Data Science Career Mentor on topmate.io. With a passion for fostering AI talent and creating innovative solutions, they are motivated by the rapidly evolving AI landscape and its potential to drive meaningful change across industries.

Location

Bengaluru

AI stack

claude code

claude

codex

cursor

Work experience

AI/ML - Lead at Kapture CX, Jun 2026 - Present · 5 mos, Bengaluru

Data Science Career Mentor at topmate.io, Jul 2024 - Present · 2 yrs 4 mos, Remote, I just want to do my part for the amazing, growing AI community. With the democratizing of myriads of AI technologies, concepts, and algorithms, it is overwhelming to decide what and where to focus when starting your career or maybe progressing to a better role in AI. I believe that I can help you in finding and "fine-tuning" some key aspects that will be useful to you for nurturing or starting a great career in Data Science. So what are you waiting for? Let's talk!

Lead Data Scientist at Oriserve, Jun 2024 - Jun 2026 · 2 yrs 1 mo, Noida, OriTTS – India's First State-of-the-Art Speech Generation Model Led the development of OriTTS, a cutting-edge TTS system with voice cloning, multi-speaker synthesis, emotion-based speech generation, cross-lingual, and transliterated TTS. Built a high-performance inference pipeline capable of real-time streaming and one-shot generation over both REST API and WebSocket. Implemented advanced acceleration techniques including DeepSpeed, FlashAttention-2, and in-flight batched streaming—achieving 2.5× throughput gains per server and sub-100ms first-byte latency. Designed a fully distributed training pipeline using PyTorch Lightning, reducing model training time from 1–2 months (on 8×A100) to just 1 week (2×4090) or 1.5 weeks (A6000). OriSTT – ASR Tailored for Indian Contexts Developed fine-tuned ASR models using custom data curation and novel training techniques. Results: lower WER, better accent handling, strong noise robustness, and direct Hindi-to-Hinglish transcription for seamless LLM integration. Deployed across multiple backends—vLLM Whisper, Whisper.cpp, Faster-Whisper, and optimized FlashAttention variants. AI Team Leadership Trained and managed a team of 9 (7 junior devs + 2 interns). Focused on autonomy, strengths-based assignments, continuous feedback, and blended responsibilities (research + dev + grunt work). Infrastructure Engineering Built scalable infra: JarvisLabs & E2E cloud setup, AI server orchestration with Nginx (load balancing), Supervisord, Uvicorn, latency logging, live metrics, and Slack alerts. Also created custom WebSocket load-testing for production. Delivered end-to-end HR voicebot with OriTTS, OriSTT, and LLMs.

Senior Data Scientist at Oriserve, Jun 2023 - Jun 2024 · 1 yr 1 mo, Noida, Email Bot(Vodafone): Implemented RoBERTA architecture for fine-tuning and context comprehension. Improved bot handling percentage by +30% Voicegenie.ai: Spearheaded the development of Voicegenie(Website: https://voicegenie.ai/), a groundbreaking Voice AI solution renowned for its generative capabilities. My involvement spanned from audio processing and optimization library integration to architecting the complex speech generation framework that underpins Voicegenie's success. Voicegenie UseCase Development: 1) Lead collection strategies for Pristyn-Care 2) Flight rescheduling solutions for Air India 3) Vodafone retention Bot 4) Debt Collection for L&T Financial Services 5) JARVIS Survey collection bot 6) Bajaj Allianz Life Insurance Outbound Voicebot 7) Tradejini Inbound Voicebot R&D: 1) Implementation of in-house generative models such as LLAMA, MPT, FALCON etc. at scale 2) Real-time voice cloning and Transliterated speech generation 3) Emotion-based Text-to-Speech model 4) In-flight batching implementation of LLMs for reducing latency and improving concurrency 5) Dynamic On-loading and Off-loading of CUDA models for efficient Memory Management and server cost reduction 6) Finetuning(full weights and PEFT) implementation of various DL model currently used for better maintenance of ML production lifecycle

Data Scientist at Oriserve, Jun 2021 - Jun 2023 · 2 yrs 1 mo, Noida, Uttar Pradesh, India, ORI is an end-to-end provider of conversational AI-powered chatbots, with the sole intention of helping enterprises make the customer experience count. Ori is focused on automating the customer journey which it does through its connected, multilingual, omnichannel, self-service platform, made by applying design thinking principles and AI. Our cognitive AI solutions provide a non-intrusive customer experience for Sales, Marketing, Support & Engagement across IoT devices, sensors, web, app, social media & messaging platforms. Our Products: 1. Chatbot: Voice + Text Compatible, Hybrid (Live + AI), Multiple Context Switching, Omni-channel along with Recovery Based Mechanisms. 2. Email Based Responder Bots: Automate Email Support with Resolutions Handed Without Manual Intervention. 3. Customer Service Agent Support Bots: Assist Agents with Prompts & Queues During Customer Conversations Over Chat + Calls. 4. Voicebot: Fully speech-based end-to-end AI voice assistant with ASR capabilites (transcription and translation, language detection, accent handling), along with realtime Text-to-Speech response back to the user, supported through both Web and cloud telephony interface.

Data Scientist at VMock, Dec 2020 - May 2021 · 6 mos, Gurugram, Completed intensive training across classical ML, deep learning, and NLP (regression, decision trees, random forests, ANNs, CNNs/RNNs/LSTMs, Word2Vec, LDA, N-grams), alongside Python backend development (Flask, JWT auth, Celery, TensorFlow Serving), building and securing an API for CSV data upload and updates. Resume Section Mapping (NLP Project) Built an NLP pipeline to automatically segment resume text into structured sections for Vmock's resume parser: Designed a Labelled LDA model to generate topic-word and word-document distribution matrices for section classification Improved on the baseline CBOW model by incorporating N-grams (bi- and tri-grams), lifting classification accuracy Enhanced feature representation using spaCy-derived noun chunks, adjectives, and verb tokens Ran systematic error analysis, optimizing the confusion matrix to refine section-bucket accuracy Tools: Python, spaCy, NLTK, Tomotopy, regex, k-Fold cross-validation

Student Researcher at RBG Labs, IIT Madras, Jul 2019 - May 2020 · 11 mos, Tamil Nadu, India, Title: Data Pre-Processing and Predictive Modelling for Road Accident Severity Prediction Guide: Dr. Venkatesh Balasubramanian, Engineering Design Department Worked on road safety analytics, predicting accident severity and identifying key contributing factors using Chennai's Road Accident Database Management System (RADMS). The project involved detecting and correcting data bias and class imbalance, applying appropriate data encoding techniques, and training multiple machine learning models to identify the best-performing approach for accuracy and generalizability. Used cross-validation and data visualization techniques to validate results and derive interpretable insights. Tools used: Python, R, ADASYN, SMOTE, k-NN, Decision Tree Classifier (DTC), K-fold cross-validation

Education

Dual Degree (Bachlors+Masters), Engineering Design, Indian Institute of Technology, Madras

Secondary and Higher Secondary Education, Jaipuria Vidyalaya