Headline

Machine Learning Engineer | Data Scientist | Speech AI (ASR/TTS) + LLM Agent Systems

About

I build the listening half of voice AI — the ASR and agent layer that has to work before a voice bot can say anything useful. At ConvoZen.ai (NoBroker) I train multilingual Conformer ASR across 9 Indian languages on 10,000+ hours and run it in production on Triton Inference Server and NVIDIA Riva at sub-second transcription. Around the model sits the harder half: a multi-session conversational harness with MCP tool-calling, multi-LLM routing, and shared state across voice, chat and email. Before this at Salesken.ai I fine-tuned NeMo's cache-aware FastConformer Hybrid into a streaming Hindi ASR model with 0–1040 ms configurable look-ahead, replaced a paid third-party API with it, and cut monthly transcription cost by 20%. It's open on Hugging Face. 🔧 What I'm working on • Multilingual ASR — distributed multi-GPU training with CTC / RNNT / TDT decoders and hybrid joint-loss objectives across 9 Indian languages, served on NVIDIA Riva • Agentic conversation systems — MCP tool-calling, multi-LLM routing with fallbacks, shared state across channels, 80% query-resolution rate across enterprise deployments • Retrieval & memory — agentic chunking that pushed Recall@1 from 63% to 80% and MRR from 0.73 to 0.87; rolling-window summarization with hybrid retrieval cutting token spend 30–40% without losing coherence • RL post-training (nights and weekends) — a POMDP clinical-reasoning environment where I post-trained Gemma-3-4B into a sequential diagnostician via SFT plus a custom GRPO loop, with a 5-rubric reward built to resist reward hacking. 6 checkpoints and 2 datasets open-sourced, OpenEnv-compatible. 🧠 Stack I work in Speech & ASR: Conformer / FastConformer, CTC / RNNT / TDT, NeMo, Whisper, speaker diarization, VAD Serving: NVIDIA Riva, Triton, TensorRT-LLM, vLLM, ONNX, Docker, Kubernetes LLM & Agents: MCP, multi-LLM orchestration, RAG, hybrid retrieval, vector DBs, PEFT, TRL / Unsloth Systems: Python, FastAPI, Kafka, PostgreSQL, MongoDB, Azure, Prometheus 🧭 How I work I'd rather read the paper and understand the tensor shapes than wire together someone else's abstraction. I care about what a system does at p99 under real traffic, not what it does in a demo. Most of the interesting work in voice AI lives in the gap between "the model is accurate" and "the conversation felt fine" — and that gap almost never closes with a better checkpoint. Open to conversations on speech AI, Indic ASR, voice agent infrastructure, and Inference Optimization.

Location

Bengaluru

AI stack

claude

hugging face

mcp

notebooklm

notion

vLLM

Work experience

Machine Learning Engineer at NoBroker.com, Jul 2025 - Present · 1 yr 4 mos, Trained multilingual ASR models on 10,000+ hours across 9 Indian languages using distributed multi-GPU training. Deployed multilingual ASR on NVIDIA Riva, delivering sub-second transcription and reducing production inference cost. Built a Memory Management Service using rolling-window summarization and hybrid retrieval, cutting token usage by 30–40% while preserving conversational coherence across long sessions. Engineered a production, multi-session conversational-AI harness with MCP tool-calling, multi-LLM routing, and shared state management across voice, chat, and email, achieving an 80% query-resolution rate across enterprise deployments. Boosted Recall@1 from 63% to 80% and MRR from 0.73 to 0.87 on an internal retrieval benchmark with an agentic chunking pipeline for RAG.

Machine Learning Engineer at Salesken, Apr 2024 - Jul 2025 · 1 y 3 mo, On-site, • Built a Retrieval-Augmented Generation (RAG) system to store and retrieve organizational data, enabling personalized, data-driven solutions and improving query resolution speed by 20%. • Developed a PII masking solution for identifying and protecting sensitive data using advanced techniques like pseudonymization, anonymization, tokenization, and truncation, ensuring compliance with data privacy regulations. • Engineered a proactive daily alert system for Slack to monitor ASR performance metrics, achieving a 30% reduction in system downtime and enhancing issue resolution processes. • Optimized and deployed the Whisper-ASR system on the Triton Inference Server, reducing latency by 25% and enhancing scalability for real-time applications. • Designed and implemented a Kafka Controller to manage message workflows efficiently, reduce data duplication by 15%, and ensure idempotency, improving system reliability and performance.

ML Intern at Salesken, Jan 2024 - Mar 2024 · 3 mos, Bengaluru, Karnataka, India, Raised ASR transcription accuracy by 10% with speaker diarization and voice-activity-detection modules, and shipped Emotion Detection models on Triton.

Virtual Internship in Generative AI at Open Weaver, Jun 2023 - Jun 2023 · 1 mo, Remote, • Developed a customized interface for Chat-GPT, acquiring hands-on experience in seamlessly integrating AI language models. • Utilized Open AI's Whisper and DALL-E tools to convert audio into visually generated images. • Applied deep learning algorithms and computer vision methods to detect and mark Deep-Fake images.

Education

Master of Science - MS, Data Science, Christ University, Bangalore

Bachelor of Science, Data Science, Gujarat University

Public links

github: https://github.com/sneh2909

sneh2909.github.io: https://sneh2909.github.io/Portfolio/