Headline

ML Engineer | Data Scientist @ convozen.ai | NIT Surat | Speech AI | Text-to-speech | Agentic AI | System Design

About

ML Engineer building production speech and conversational AI systems. Currently at Convozen.ai (NoBroker) shipping voicebot infrastructure that handles 1.5M+ voice calls/day. My focus is the unglamorous-but-decisive layer of voice AI: streaming pipelines, latency budgets, agentic tool-calling, and the messy work of making LLMs feel human in real-time conversation. On the side, I'm building out the Gujarati speech AI ecosystem — I've open-sourced 4 first-of-their-kind Gujarati TTS models on Hugging Face (F5-TTS, XTTS-v2, Chatterbox, Piper) trained on a 150-hour dataset I curated myself. 🔧 What I'm working on Voicebot infrastructure at scale — websocket pipelines, chunked LLM streaming, dynamic filler injection, multi-provider LLM routing with fallback handlers Agentic systems — custom action frameworks, parallel tool execution, agentic chunking for RAG (pushed Recall@1 from 63% → 80% on our production pipeline) Indic speech AI — TTS finetuning (flow matching, DiT internals), ASR benchmarking, end-to-end data curation pipelines for low-resource languages 🧠 Stack I work in Speech & Audio: F5-TTS, XTTS-v2, Piper, Chatterbox, Whisper, MFA, WebRTC VAD, librosa LLM & Agents: LangGraph, MCP, Qdrant, FAISS, hybrid search, neural rerankers, flow matching Systems: Python (AsyncIO), FastAPI, WebSockets, Docker, GCP (Cloud Run, GCS), PostgreSQL 🧭 How I work I like building things from first principles — reading the paper, understanding the tensor shapes, then shipping. I care more about whether a system actually works under production load than whether it demos well. Most interesting problems live at the boundary between "the model can do this" and "the user experiences this as working." Open to conversations on speech AI, Indic NLP, voicebot infrastructure, and anything at the intersection of research and production. 📍 Bengaluru | 🤗 huggingface.co/Arjun4707 | 💻 github.com/BhammarArjun

Location

Bengaluru

AI stack

cursor

opencode

Work experience

Machine Learning Engineer at NoBroker.com, Jul 2025 - Present · 1 yr 2 mos, Bengaluru, Karnataka, India, Building in-house TTS model for Indic use cases . at <$3 per million tokens and <200 ms first-chunk latency, benchmarked across T4, L4, A100, RTX 6000 Pro, H100, and H200 GPUs Building production voicebot infrastructure for an enterprise Conversational AI platform handling 1.5M+ voice calls/day. Key work: Engineered an ultra-low-latency speech pipeline using websockets and chunked LLM streaming — reduced end-to-end voicebot latency by 25% Designed an asynchronous voice filler injection system that masks background processing spikes with conversational fillers ("uh-huh", "hmm"), driving perceived latency to near-zero Built a custom Agentic Action Framework from scratch — autonomous tool-calling with parallel API execution driven by real-time LLM schemas Developed a modular multi-provider LLM integration layer (Google GenAI, OpenAI, custom REST) with latency-profile-driven routing and automatic failover Architected an Agentic Chunking pipeline for our RAG stack — Retrieval Recall@1 jumped from 63% to 80%, MRR from 0.73 to 0.87, complex table/text accuracy up to 55%

Data Scientist at Finsire, Mar 2024 - Jun 2025 · 1 yr 4 mos, Pune District, Maharashtra, India, Data Scientist working on lending infrastructure for India's first asset-backed credit platform — building the data systems behind instant underwriting against mutual funds, vehicles, and bank-statement-derived risk profiles. Key work: Automated NAV Pipeline: Designed and deployed an automated daily Mutual Fund NAV update and Loan-to-Value (LTV) calculation pipeline. Cut processing time from 4 hours to 1 minute — a 240x speedup — eliminating manual reporting bottlenecks for the LAMF (Loan Against Mutual Funds) product. Bank Statement Analyzer: Built a high-throughput parsing and analysis system processing raw transaction JSON through deterministic and LLM-based logic layers. Achieved 85% transaction categorization accuracy and extracted deep risk profiles — loans, investments, recurring debits/credits — for instant underwriting decisions. IDV Valuation Engine: Engineered a vehicle pricing framework leveraging 15+ years of historical market data and custom depreciation formulas to automate Insurance Declared Value (IDV) assessments for collateral-backed lending programs. Stack: Python, PostgreSQL, FastAPI, GCP (Cloud Run, GCS), LLM integration for unstructured data extraction.

PGTE at Bajaj Auto Ltd, Aug 2023 - Dec 2023 · 5 mos, Spearheaded the development and optimization of SOPs for over 10 production activities, providing comprehensive guidance to new trainees and enhancing operational efficiency. Implemented motion economy principles to streamline standard work procedures for 16 box-filling workers, resulting in increased efficiency and reduced fatigue levels. Conducted manpower and work content analysis for the engine packing conveyor, leading to a direct enhancement in production capacity and operational effectiveness. Collaborated with cross-functional teams to identify process improvement opportunities and drive continuous enhancements in production work-flows. Led training sessions for new trainee & workers, focusing on best practices, safety protocols, and quality standards to ensure a high level of operational excellence. Participated in quality control initiatives, ensuring adherence to industry standards and regulatory requirements for all production activities. Executed data-driven decision-making processes to enhance operational efficiency and drive continuous improvement initiatives.

Project Intern at Tata Motors, Jan 2020 - Mar 2020 · 3 mos, Sanand, Gujarat, India

Summer Internship at Bosch Rexroth, May 2019 - May 2019 · 1 mo, Sanand, Gujarat, India

Education

Master's degree, Manufacturing Engineering, Sardar Vallabhbhai National Institute of Technology, Surat

Bachelor's degree, Mechanical Engineering, Nirma University