Kshitij Sharma

Lead AI/ML Engineer & Architect

• Architected and built a multi-camera retail vision platform for live store environments, integrating customer analytics with Petpooja, Zoho, and delivery data through Django/FastAPI microservices and REST APIs.
• Engineered real-time detection with YOLO12s, Kalman/Hungarian tracking, and cross-camera re-identification and facial recognition with OSNet/ArcFace and FAISS indexing, sustaining 65–85 ms median per-camera latency on 2.7K feeds.
• Developed homography-based world-space fusion to maintain identity continuity across overlapping camera views and derive dwell time, table occupancy, customer retention, movement flows, and zone activity.
• Built a multi-tenant restaurant AI platform for web and Telegram using LLMs, LangChain, RAG, and FAISS for menu discovery, cart management, ordering, and order tracking. Integrated persistent context, semantic caching, POS systems, human-agent takeover, and natural-language sales analytics.
• Deployed Dockerized edge/cloud services using FastAPI/Django, PostgreSQL, GPU infrastructure, CI/CD, Cloudflare security, and production monitoring.

Lead Data Scientist · Consultant, remote

• Developed context-aware vehicle intelligence using probabilistic suggestions and self-learning routines for comfort settings, heating, onboard applications, and location-based suggestions.
• Applied Bayesian inference and variational modeling to user behavior and in-car action prediction, supporting adaptive suggestions and personalized driver experiences.
• Led causal inference and machine learning initiatives to optimize vehicle climate control, EV charging strategies, and other operational prediction problems.
• Owned the ML lifecycle from Python research to embedded C++ implementation, authoring production software deployed onto 2026 Mercedes-Benz production vehicle models.

Lead Data Scientist

• Built hyper-personalized playlist models serving 50M+ users through music embedding spaces, nonparametric clustering, and recommendation techniques.
• Designed and deployed a learning-to-rank model for personalized search, improving NDCG with deep NLP architectures integrated with Elasticsearch.
• Engineered Python, TensorFlow, and Spark pipelines to learn song embeddings from audio features (MFCC) and user consumption patterns, powering similarity and ranking engines.

Machine Learning Engineer

• Developed vehicle prediction models using Bayesian inference and Bayesian nonparametric methods, including mixture-model approaches to location clustering and automation.
• Designed, implemented, and tested ML models in C++ for production deployment within vehicle head units.
• Built causal graphs and counterfactual models to distinguish low-value head-unit applications from those underused because of availability or placement, informing application discontinuation decisions.

Software Engineer, Research Intern

• Developed backend features for SAS Graph Template Language in a C++ graph framework, including new SAS procedures and visualization functionality.
• Migrated graphing procedures and functionality from Java to the newer C++ platform.

Software Engineer

• Developed backend services for a multi-channel Java commerce platform, with primary ownership of order capture and associated commerce functionality.
• Integrated payment processing, shopping baskets, and web services for end-to-end transactions across channels.
• Designed an in-memory cache for high-speed inventory transactions to reduce repeated database and backend lookups.