Headline

ML & AI Engineer | Ex-INRIA | Patent Holder

About

I build AI systems that work in the real world — at scale, under pressure, and in production. Over the past 4+ years, I've worked across the full AI stack: from deploying sensor-fusion models on live autonomous vehicles at INRIA (Grenoble) to architecting Voice AI agents that now autonomously manage a ₹2 Billion customer portfolio at 4x lower cost than human agents — 1,000+ calls every day. My work sits at a rare intersection: I'm both a builder and a researcher. I've publications, hold an Indian patent, and have shipped production systems that replaced real human workflows. That combination — research rigour applied to real-world constraints — is what I bring to every problem. Currently at WeRize, I'm deep in the Voice AI space: building multilingual agents, benchmarking STT/TTS/LLM vendors across the Indian market, and designing Speech Analytics pipelines that deliver structured intelligence from every customer call. Before that, 3+ years at INRIA in France working on autonomous vehicle perception — real-time 3D object detection, event camera sensor fusion, and spatio-temporal scene forecasting on moving vehicles. If you're working on hard AI problems — in voice, autonomy, or LLM systems — I'd love to connect.

Location

Bengaluru

AI stack

antigravity

codex

Work experience

Artificial Intelligence Engineer at WeRize, Sep 2025 - Present · 1 yr, Bengaluru, Karnataka, India, • Architected and deployed multilingual LLM-powered Voice AI agents, automating 80% of outbound collections calls across a INR 2B+ portfolio and handling 1,000+ daily calls, achieving 4× cost reduction vs human agents through modular multi-agent orchestration. • Led benchmarking of 8+ Voice AI platforms, evaluating ASR accuracy (Indian accents), latency, and cost, driving production vendor selection for enterprise-scale deployment. • Built a scalable AWS-based speech analytics platform processing thousands of calls/day, integrating ASR, speaker diarization (95% accuracy), sentiment analysis, and LLM-driven QA scoring — eliminating manual QA workflows and improving audit efficiency by 90%

Prompt Engineer at TELUS Digital AI Data Solutions, May 2025 - Jun 2025 · 2 mos, Remote, Crafted 100+ production-ready prompts for fine-tuning domain-specific LLMs, using standardized formats to encode complex multimodal inputs (context + images) for real-world applications.

AI Engineer at Inria, Apr 2021 - Oct 2024 · 3 yrs 7 mos, Grenoble, Auvergne-Rhône-Alpes, France, [ Writing Technical Project Proposals and Grants ] • Project 1: Developed spatio-temporal prediction network (LSTMs) pipeline for accurately forecasting future occupancy in dynamic urban environments, predicting 3-sec horizon, and releasing valuable occupancy grid dataset for research advancement. (IV’22) • Project 2: Designed a robust sensor fusion model combining eventbased and frame-based cameras, to enhance object detection in challenging conditions, surpassing frame-based and event-based detections, with over 30% improvement in robustness to image corruptions, validated on the publicly released DSEC dataset. (ICRA’22) • Project 3: Deployed a real-time 3D object detection and tracking module in ROS, achieving performance at 10FPS for vehicle demonstration purposes. • Project 4: Leveraged generative models and model-free deep RL algorithms to excel in complex urban traffic scenarios using Carla simulator, with a focus on applying Variational Auto-Encoder (VAE) for dimensionality reduction. • Project 5: Designing formal validation methodologies for probabilistic perception and prediction methods, and analyzing real experiments recorded data - PRISSMA project. • Project 6: Fine tuning LLMs for Autonomous driving applications

Artificial Intelligence Engineer at Inria, Apr 2021 - Oct 2024 · 3 yrs 7 mos, • Real-Time 3D Perception (Vehicle Deployment): Deployed a TensorRT-quantized PointPillars model for real-time 3D object detection and multi-object tracking on a live moving vehicle at 15–20 km/h, achieving 15.15 FPS — eliminating critical latency gaps that caused missed object detections at speed. • Multimodal Sensor Fusion: Developed a novel RGB + event camera fusion network (RetinaNet backbone) trained on 131K+ annotated bounding boxes. Achieved 68.7% relative performance and 30% robustness gain over RGB-only baselines across 15 real-world corruption scenarios. (ICRA’22) paper code, video • Spatio-Temporal Scene Forecasting: Built an LSTM-based pipeline forecasting 3-sec future occupancy grids in dynamic urban environments from nuScenes data. Achieved 82.9% F1-score over 1,200+ real-world test sequences, surpassing prior state-of-the-art. (IV’22) paper code, video • Multimodal Data Engineering: Curated and synchronized LiDAR, and RGB camera data streams; analyzed collision edge-case scenarios to validate the full perception stack.

Student Assistant at Max Planck Institute for Intelligent Systems, Nov 2020 - Feb 2021 · 4 mos, Tübingen, Baden-Württemberg, Germany, Working with Prof. Aamir Ahmad Robot Perception Group - Perceiving Systems Department Topic: Development of outdoor navigating robot

Autonomous Systems at EIT Digital, Aug 2018 - Feb 2021 · 2 yrs 7 mos

Education

Master's degree, Autonomous Systems, Technische Universität Berlin

EIT Digital Summer School, Data Driven Manufacturing for Industry 4.0, Alma Mater Studiorum – Università di Bologna

Public links

github: https://github.com/ksm26