Resume
Sanket Waghmare
Lead AI/ML Engineer | Computer Vision, GenAI & Edge AI | Production AI Systems
Hyderabad, India · sanketw1012@gmail.com · LinkedIn · GitHub
AI/ML engineer with 7+ years spanning computer vision, classical ML, generative AI/LLMs, agentic systems, edge computing, and industrial/physical AI. Built and deployed production CV pipelines on constrained edge hardware (NVIDIA Jetson, TensorRT, DeepStream), classical ML systems for marketplace forecasting and recommendation, and enterprise GenAI platforms: multi-tenant RAG, natural-language-to-SQL, and agentic SDLC automation. A founding technical contributor to a 0→1 hardware-AI program that proved feasibility for a physical Deposit Return System later scaled into a ~50-person organization and a 500-machine manufacturing program. Pursuing PhD research in computer-vision image restoration (Single Image Reflection Removal) alongside the role. Provides technical leadership across AI/ML, robotics, embedded, and software engineering, while remaining hands-on in system design, modeling, and deployment.
Core expertise
- AI / ML
- Deep learning, time-series forecasting, recommendation systems, knowledge distillation, open-set recognition, experiment design, model evaluation, production ML
- Computer vision
- Detection, instance and semantic segmentation, tracking, image restoration, visual similarity search, industrial inspection, real-time inference
- GenAI / LLM
- RAG, agentic AI, tool calling, natural-language-to-SQL, document AI, knowledge systems
- Frameworks
- LangChain, LangGraph, LlamaIndex, Qdrant, ChromaDB, FAISS, FastEmbed
- Edge AI
- NVIDIA Jetson Nano and Xavier NX, Raspberry Pi 4, TensorRT, DeepStream, ONNX, OpenCV, industrial cameras
- Software and cloud
- Python, PyTorch, TensorFlow, FastAPI, Flask, MySQL, SQLAlchemy, Pandas, PySpark, AWS, GCP Vertex AI, Azure Databricks, MLflow
Experience
Recykal · Hyderabad, India · Aug 2021 – Present
Associate Director – Technology · Apr 2026 – Present
Technical lead for AI/ML, computer vision and applied AI. Owns AI architecture across the Deposit Return System program and the enterprise AI platform, with technical direction of a ~20-person cross-functional group and ~10 direct reports.
- ReTrieve, an enterprise agentic knowledge platform: multi-tenant hybrid retrieval (dense and sparse, fused with Reciprocal Rank Fusion in Qdrant) with access control enforced at retrieval time; a Gemini function-calling agent that asks for user confirmation before widening its search; domain-aware federated retrieval; capabilities exposed over MCP with OAuth 2.1/PKCE. In use by ~30 internal users over 10,000+ document pages.
- Reflow, a multi-channel communication and operations intelligence system: a three-service architecture unifying WhatsApp, chat exports and Gmail into one event model; OCR, classification and action extraction with entity-specific prompts; shipment-based grouping across channels; a knowledge graph; and direct linking of WhatsApp-sourced documents to transactions, which removed manual reconciliation for those documents. 100,000+ messages and emails processed.
- Agentic SDLC automation with LangGraph and LangChain: requirements and Jira context turned into a connected chain of PRD, HLD, LLD and test plans, with human-in-the-loop checkpoints before writing back to Jira or GitLab. Process tooling, not autonomous development.
- AI adoption: identified document reconciliation and cross-channel vendor communication as recurring automation opportunities across functions, directed platform-level solutions instead of one-off tools, and moved prototypes into production services with consistent authentication and access control.
Senior Engineering Manager – AI · Apr 2025 – Mar 2026
Technical lead for AI/ML architecture across GenAI, computer vision, MLOps and enterprise analytics.
- EPR Conversational Analytics: natural-language-to-SQL over enterprise data with schema retrieval (ChromaDB), a customised Vanna-style approach, pre-execution validation against double counting in many-to-many relationships, a feedback loop from analyst-approved corrections, and offline evaluation with LLM-as-judge. ~98% SQL accuracy on an internal 100-question benchmark; ~10 active users.
- Domain-specific RAG services (CircularGPT, DRS Bot): multiple independently indexed Qdrant knowledge domains with routing to a single domain, falling back to parallel federated retrieval when the domain is ambiguous.
- Document AI: classification-then-extraction pipeline for invoices, weighbridge documents, e-way bills and logistics documents, with page-level splitting of mixed scans. In production at ~50,000 documents per month; ~7x faster internal processing; extraction accuracy from ~60% to ~90% across three technology generations (rules and OCR, then Donut, LayoutLM and PaddleOCR, then Gemini-based extraction).
- Compact CLIP distillation for open-set foreign-object rejection on a Jetson Nano: a ~1.2M-parameter MobileNetV3-Small student, frozen text anchors (~20 KB), negation-aware training and student-space threshold recalibration. On a 46,123-image held-out set: 91.91% foreign-object recall and 99.77% precision (an experimental result, not a claim about all production machines).
Engineering Manager – AI · Apr 2024 – Mar 2025
Technical lead for production computer vision and edge AI.
- Architected the production computer-vision system for Reverse Vending Machines: material, brand, SKU, colour and quality decisions, and foreign-object detection, across ~3,000 brands and ~500 SKUs. YOLO-family models optimised for NVIDIA Jetson with TensorRT, DeepStream and ONNX. ~98% true-negative performance at 20–30 ms inference, across ~350 machines in production.
- Conveyor sorting AI (Reklaim Line): extended the same detection, segmentation and tracking approach to high-throughput conveyors. Two production lines, ~100,000 items per day at ~98% accuracy, the subject of granted Indian Patent No. 592451.
- PET bottle contamination grading: a deep-learning plus classical-CV hybrid producing an interpretable Dirtiness Score, deployed on Jetson Xavier with an ESP32-driven ejector. On 10,442 field images: ~93.36% overall accuracy at ~2–20 ms per image, running at 120–240 items per minute.
- Contributed to a fraud-intelligence proof of concept (proof-of-concept stage only, not in production) and to operational analytics for the Deposit Return System program.
Senior Software Engineer – AI · Apr 2023 – Mar 2024
Founding technical contributor to the Retearn / Deposit Return System hardware-AI initiative, alongside SmartSkan productionisation and marketplace machine learning.
- Retearn Technologies (0→1): proved the initial technical concept, built early AI prototypes and evaluated feasibility of real-time computer vision on embedded hardware. The initiative later scaled into an organisation of ~50 people and a 500-machine manufacturing program. This was an outcome the early technical work helped enable, not a personal management claim.
- B2B marketplace ML: a univariate LSTM price forecaster, supply and demand date and quantity models, a logistics price predictor and a seller–buyer recommendation engine.
- MLOps on AWS EC2 and MLflow: experiment tracking, versioning and reproducibility across the vision and classical-ML workstreams.
Software Engineer – AI · Aug 2021 – Mar 2023
Individual contributor with core ownership of the SmartSkan computer-vision system from first build to first production deployment.
- SmartSkan: an end-to-end waste-vision pipeline built with IoT image acquisition, a curated dataset of 3M+ images in ~200 hierarchical categories, Mask R-CNN instance segmentation with DeepSORT tracking, deployed on GCP Vertex AI, developed in the context of a collaboration with Google's sustainability team associated with CircularNet. Historical internal results above 90% for the Plastic and Metal categories (not a current production figure).
- Visual Quality Check: a contamination-segmentation pipeline using colour-space bootstrapping, expert-refined masks and a custom U-Net, estimating contamination percentage in a batch.
Cardinal Intellectual Property · Chicago, USA · May 2019 – Jun 2021
Computer Vision Engineer · May 2019 – Jun 2021
Patent image search engine.
- Designed and productionised visual patent-image search: a retrained EfficientNet feature extractor for patent drawings, FAISS similarity search, a Flask REST API with a web interface, automated MS-SQL pipelines, and feature databases maintained in Azure Databricks. Supports whole-image and cropped-region queries.
Patents
- A Method and System for Automated Waste Counting, Identification, Classification, and Sorting with Artificial IntelligenceIndian Patent No. 592451, granted by the Indian Patent Office on 17 June 2026. Relates to the Reklaim Line sorting system.
- System and Method for Managing Segregation and Collection of Waste Using Artificial IntelligencePatent application 202241053571, published in Journal 41/2022. Application stage only, not granted.
Education
- PhD, Computer Vision & Pattern Recognition (in progress, 2022 – ongoing)Indian Institute of Technology Ropar. Research on single image reflection removal; not yet published.
- M.S., Electrical & Computer Engineering (2017 – 2019)University of Illinois Chicago. Computer vision and machine learning specialisation.
- B.E., Instrumentation Engineering (2011 – 2015)Mumbai University.
Certifications
- Deep Learning Specialization, deeplearning.ai (2018)
- Machine Learning, Stanford University (2018)