I design and ship production-grade AI systems — real-time voice agents, LLM pipelines, and enterprise NLP tools. I care deeply about latency, scalability, and systems that actually work in the real world — not just demos.
Currently focused on Voice AI: low-latency speech-to-speech pipelines combining ASR, LLMs, and TTS for real-world conversational agents at scale.
🎙️ Voice AI Engineer — TeleCMI Communications Sept 2025 – Present
🔴 Real-Time Voice Agent System
- Designed a real-time speech-to-speech voice agent achieving <600ms end-to-end latency under concurrent load
- Cut pipeline latency 1.3s → <600ms (~54% reduction) via model optimization & pipeline parallelism on a single GPU
- Integrated Ultravox (ASR+LLM) · Qwen (reasoning) · VibeVoice (TTS) with RAG-based dynamic retrieval
🟠 Post-Call Analytics System
- Processes 400+ hours of calls/day across 2 languages with ~95% transcription accuracy + automatic language detection
- LLM-based analysis extracts customer pain points, intents, and product insights from call transcripts
🟡 End-to-End Voice Data Processing Pipeline
- Audio ingestion via Apache Pulsar → denoising → speaker diarization → transcription, all production-grade
- Outputs structured JSON metadata + segmented audio to AWS S3 for analytics and model training
🤖 AI/ML Engineer — Motherson Technology Services Limited June 2024 – Sept 2025
🔵 Confidential Knowledge Chatbot — Motherson Group
- Built a secure RAG system over 60+ GB of confidential documents across 100+ companies — <2.5s query latency
- Centralized semantic search across the entire group, reducing redundant work and surfacing prior knowledge
🟣 M&A Contract Analysis
- Automated comparison of 4+ legal contracts (200–300 pages each), saving ~1 month of due diligence time
- LLM-based structured analysis of high-volume M&A agreements
🟢 47×47 Multilingual Enterprise Translator
- Deployed a translation system across 47 source → 47 target languages using OpenAI + Helsinki-NLP/Opus
- Supports
.txt,.docx,.pdfwith XML-based table structure preservation
|
Modular Real-Time Voice Agent Orchestrator-based architecture that decouples ASR · LLM · TTS instead of relying on monolithic models like Ultravox. WebSocket integration with Pipecat, supporting Qwen + AI4Bharat Indic Conformer with extensibility for LLaMA, Mistral, and more.
|
🔊 DieTraAudio Processing Pipeline Denoising → Diarization → Transcription using DeepFilterNet, NVIDIA NeMo, and Mistral Voxtral. Runs on 10–12GB VRAM vs traditional 80GB+ — high-quality audio processing on consumer hardware.
|
|
7 Attention Mechanisms Clean implementations from vanilla self-attention to Flash Attention and Multi-Latent Attention, with complexity analysis. A deep dive into modern LLM internals.
|
AI-Powered MoM Generator Whisper + GPT-3.5 + PyAnnote diarization with a React frontend. Auto-generates professional meeting minutes with full speaker attribution.
|
|
RAG Conversational AI Multi-turn context-aware chatbot with document retrieval and memory using LangChain.
|
GPS-Free Positioning · Guide: Dr. Amitangshu Pal Sensor fusion (accelerometer + compass + ML) achieving ±20–30m accuracy over 4–6 km. Works with standard smartphone sensors.
|
"Building real-time AI systems, one model at a time." 🎙️