---
title: "Avinash Benki - Sr Staff AI Engineer in Bengaluru | Tal"
description: "Meet Avinash Benki, Sr Staff AI Engineer | Multimodal Foundational AI | ASR · LLM · TTS · Conversational AI | Indic Language AI | 7+ yrs Building Real-World NLP at Scale based in Bengaluru."
image: "https://tal.club/meet/avinash-benki/share-image"
---

# Avinash Benki on Tal

## Headline

Sr Staff AI Engineer \| Multimodal Foundational AI \| ASR · LLM · TTS · Conversational AI \| Indic Language AI \| 7\+ yrs Building Real\-World NLP at Scale

## About

What I Do I build production\-grade NLP and voice AI systems — from multimodal Audio LLMs to low\-latency speech pipelines\. With 6\+ years at gnani\.ai, I've worked across the full stack of conversational AI: from training TTS models from scratch to deploying RAG\-based knowledge systems serving thousands of enterprise users\. Recent Highlight: Contributed to gnani\.ai's Multimodal Speech Audio LLM — a model that fuses acoustic signals with language understanding to enable context\-aware, real\-time voice intelligence\. This sits at the frontier of where speech processing meets large language models\. Impact I've Delivered: \- Optimized voice bot latency to under 2 seconds for Indic languages via enhanced orchestration & inference \- Achieved 35% improvement in suggestion accuracy with DPO fine\-tuning \+ 3x cost savings vs\. OpenAI \- Built RAG knowledge system that cut bot development time by 20% \- Scaled NLP infrastructure to 30,000 concurrent outbound calls and 1,000 req/sec \- Automated EMI debt collection for TVS Credit & Bajaj Life Insurance — boosted monthly collection by 45%\+ \- Built TTS systems for Kannada and Hindi using Tacotron2 \+ WaveGlow — 900ms inference on CPU My Technical Stack \- LLMs & Fine\-tuning: vLLM · DPO · LoRA · Llama · Mistral · Cohere · OpenAI · NVIDIA NeMo \- Speech AI: ASR · TTS \(Tacotron2, WaveGlow\) · Backchanneling · Indic Language Models · LiveKit \- RAG & Agents: LangChain · Crew AI · LangGraph · OpenAI · Llama\-3\.1 · Extractive QnA \- Infra: vLLM · Docker · CUDA · Flask · Python · Multi\-GPU Training What I'm Passionate About: I believe the next breakthrough in AI isn't just smarter models — it's AI that truly listens and understands human voice in every language, dialect, and context\. India has 22\+ official languages and over a billion people who deserve world\-class voice AI\. I'm building that\.

## Location

Bengaluru

## Work experience

Sr Staff AI Engineer at gnani\.ai, Jun 2026 \- Present · 3 mos, Bangalore Urban

Lead NLP Engineer at gnani\.ai, May 2024 \- Jun 2026 · 2 yrs 2 mos, Bengaluru, \- Contributed to gnani\.ai's Multimodal Speech Audio LLM — integrated acoustic encoder representations with large language model layers to enable context\-aware, real\-time spoken language understanding at enterprise scale \[https://www\.cnbctv18\.com/technology/bengaluru\-based\-gnani\-ai\-unveils\-voice\-to\-voice\-ai\-model\-at\-india\-ai\-summit\-ws\-l\-19853427\.html\] \- Optimized voice bot end\-to\-end latency to under 2 seconds for Indic languages through enhanced orchestration pipeline and inference engine tuning \(NVIDIA NeMo \+ LiveKit\) \- Led a cross\-functional team of 5 NLP engineers to optimize call center workflows for a U\.S\. Obamacare health insurance client — significantly reduced Average Handle Time \(AHT\) and improved agent efficiency \- Designed and deployed real\-time LLM\-based agent assist system: dialog summarization, automatic note creation, and context\-aware suggestions using an open\-source LLM \- Architected a Retrieval\-Augmented Generation \(RAG\) system to streamline agent information retrieval, reducing manual lookup time for insurance domain queries \- Fine\-tuned LLM using DPO \(Direct Preference Optimization\) and deployed via vLLM — achieved 35% improvement in suggestion accuracy with 3x cost savings vs\. OpenAI API \- Drove Multilingual LLM roadmap for Indic language support across Hindi, Bengali, Kannada, and Tamil voice bot deployments

Senior NLP Engineer at gnani\.ai, Aug 2021 \- May 2024 · 2 yrs 10 mos, Bengaluru, Senior NLP Engineer – Collaborated with product managers, end clients, backend and frontend engineers, and planned deployments while driving team success in developing innovative NLP solutions\. \- Developed a hybrid LLM/NLU chatbot architecture with intelligent semantic routing, leveraging cosine similarity to route queries and achieving a 90% reduction in latency for high\-confidence intents in the insurance domain\. \- Built an end\-to\-end RAG\-based knowledge base system using OpenAI, Llama\-3\.1, and Cohere models\. Incorporated extractive QnA capabilities and efficient caching mechanisms to optimize question answering\. \- Implemented an LLM\-driven workflow co\-pilot using Crew AI, generating workflow graph structures from script or descriptive input, reducing bot development time by an estimated 20%\. \- Enhanced the user experience of LLM\-based voice bots by implementing backchanneling and filler word injection models powered by a DeBERTa\-based classifier, reducing first\-token latency from 150ms to 50ms\. \- Improved test coverage by 80% and reduced testing time by 60% for LLM chatbot testing through a lang\-graph based agentic system that automatically generates diverse multi\-lingual test cases and conversations, including edge cases\. \- Developed a campaign manager using a Python backend to automate outbound calling, scaling to more than 30,000 concurrent calls\.

Assistant System Engineer Trainee at Tata Consultancy Services, Jun 2018 \- Nov 2019 · 1 yr 6 mos, Pune Area, India, Support and Enhancement for Banking Application on Documentum

## Education

Bachelor's degree, Mechanical Engineering, Siddaganga Institute of Technology

Machine Learning Nanodegree, Information Technology, Udacity

[View the original profile](https://tal.club/meet/avinash-benki)

```json
{
  "@context": "https://schema.org",
  "@graph": [
    {
      "@type": "Person",
      "@id": "https://tal.club/meet/avinash-benki#person",
      "name": "Avinash Benki",
      "url": "https://tal.club/meet/avinash-benki",
      "image": "https://tal.club/meet/avinash-benki/profile-photo.jpg",
      "description": "Meet Avinash Benki, Sr Staff AI Engineer | Multimodal Foundational AI | ASR · LLM · TTS · Conversational AI | Indic Language AI | 7+ yrs Building Real-World NLP at Scale based in Bengaluru.",
      "jobTitle": "Sr Staff AI Engineer",
      "worksFor": {
        "@type": "Organization",
        "name": "gnani.ai"
      },
      "hasOccupation": {
        "@type": "Occupation",
        "name": "Sr Staff AI Engineer"
      },
      "homeLocation": {
        "@type": "Place",
        "name": "Bengaluru"
      },
      "alumniOf": [
        {
          "@type": "EducationalOrganization",
          "name": "Siddaganga Institute of Technology"
        },
        {
          "@type": "EducationalOrganization",
          "name": "Udacity"
        }
      ],
      "mainEntityOfPage": {
        "@id": "https://tal.club/meet/avinash-benki#webpage"
      }
    },
    {
      "@type": "ProfilePage",
      "@id": "https://tal.club/meet/avinash-benki#webpage",
      "url": "https://tal.club/meet/avinash-benki",
      "name": "Avinash Benki on Tal",
      "description": "Meet Avinash Benki, Sr Staff AI Engineer | Multimodal Foundational AI | ASR · LLM · TTS · Conversational AI | Indic Language AI | 7+ yrs Building Real-World NLP at Scale based in Bengaluru.",
      "dateModified": "2026-08-09T14:51:47.992Z",
      "isPartOf": {
        "@id": "https://tal.club/#website"
      },
      "mainEntity": {
        "@id": "https://tal.club/meet/avinash-benki#person"
      },
      "primaryImageOfPage": {
        "@type": "ImageObject",
        "url": "https://tal.club/meet/avinash-benki/profile-photo.jpg",
        "contentUrl": "https://tal.club/meet/avinash-benki/profile-photo.jpg",
        "caption": "Avinash Benki profile photo on Tal"
      },
      "inLanguage": "en-IN"
    }
  ]
}
```