Mangesh Damre

Software Engineer with 2.5+ years of industry experience in C/C++ systems programming, ML inference frameworks (ONNX Runtime, TensorRT, TFLite), and AI orchestration pipelines on Linux and Windows platforms. Hands-on experience building and optimising ML runtimes, distributed inference systems (Ray), and modular SDK-style components for heterogeneous computing targets including H100 GPU, NVIDIA Jetson (edge), and Apple M2 (on-device). Strong foundation in data structures, algorithms, object-oriented design, CMake-based build systems, and Git workflows.

Experienced debugging performance bottlenecks across system and application layers; familiar with quantisation (INT8/FP16/Q4), execution graphs, model loading pipelines, ONNX Runtime integration, and agentic AI orchestration. Eager to contribute to production ML system development for Snapdragon platforms.

Associate AI Engineer (Researcher) - Centific.ai - Hyderabad, India

(2025-04)

• ML Inference & AI Orchestration Pipeline (C++/Python/Ray): Designed and deployed a Ray-based distributed ML inference orchestration framework for automated multimodal video analysis; integrated PyTorch and ONNX Runtime models with GPU-accelerated scheduling (0.04–1.0 GPU/task allocation), reducing multi-hour processing to minutes at 90%+ accuracy. Implemented model loading, task scheduling, and runtime integration patterns directly applicable to Qualcomm Orchestrator SDK development.
• ONNX Runtime Integration & Model Optimisation: Built production inference pipelines using ONNX Runtime and TensorRT with INT8/FP16/Q4 quantisation and model compression; deployed on NVIDIA H100 (cloud) and Jetson Nano (edge) – spanning the same heterogeneous compute stack as Snapdragon QNN/SNPE targets. Validated model execution graphs and quantised model accuracy across precision levels.
• SDK-Style Modular ML Microservices (FastAPI/Python): Engineered modular, SDK-inspired ML microservice components (Whisper ASR, pyannote diarization, Presidio PII detection) with clean API contracts, unit test harnesses (37 passing), and automated QC validation pipelines – mirroring SDK component development practices.
• NVIDIA Cosmos VLM Suite – Runtime Integration & Evaluation: Integrated Cosmos-Reason1-7B, Cosmos Transfer, Cosmos 2, and Cosmos 3 VLMs into inference orchestration pipelines on H100; built custom evaluation harnesses to benchmark output quality and execution performance across model variants. Debugged CUDA library conflicts and resolved runtime integration blockers (NVCF authorisation, DGX Cloud dependencies).
• Physical AI Data Platform (C++/Python/AWS/PostgreSQL): Architected a multi-source ML data ingestion platform consolidating 7+ vendor streams (Azure Blob, AWS S3, SharePoint) into a PostgreSQL-backed discovery catalog (7 tables, incremental change detection) – demonstrating cross-platform integration and storage system design (SQL + NoSQL/S3).
• On-Device LLM Inference (C/C++, CMake, llama.cpp, GGML/GGUF): Compiled llama.cpp from source using CMake; ran 4-bit quantised (Q4 K M) 1.1B-parameter LLM fully on-device via Metal backend on Apple M2 GPU. Benchmarked real-world throughput: 985.6 tok/s prompt evaluation, 109.7 tok/s generation. Systematic analysis of quantisation vs. speed/quality trade-offs across Q2→F16 – directly analogous to Snapdragon edge inference work.
• NVIDIA TAO Toolkit – ML Model Optimisation Pipeline: Built end-to-end optimisation pipelines for Faster R-CNN, DetectNet V2, EfficientDet V2, DINO, and Grounding DINO; applied INT8/FP16 quantisation and model compression to meet accuracy/latency SLAs on Jetson edge hardware and H100 cloud.
• Computer Vision Inference Modules: Built and optimised YOLO11 object detection, Vision Transformer (ViT-Large) inference, 360 video unwarping, and VLM scene captioning modules using ONNX Runtime and TensorRT; profiled and debugged latency across system and application layers.
• VLA Training Data Platform & Robotics Pipeline: Designed end-to-end VLA training data platform – egocentric video ingestion, multimodal annotation, human-to-robot motion retargeting, HDF5 dataset packaging (LeRobot v2.1/GR00T-compatible) – managing MCAP/HDF5 data workflows and Rerun-based visualisation and QC.

Software Engineer – AI Integration - LTIMindtree - Mumbai, India

(2022-07 - 2023-08)

• Integration Pipeline Development (MuleSoft/Boomi): Built scalable integration pipelines for AI-enabled APIs using MuleSoft Anypoint Platform and Boomi Atmosphere; contributed to cross-platform data flow between cloud analytics systems and downstream ML services – including shared library integration patterns.
• ML Microservice Containerisation (Docker): Led containerisation of ML microservices using Docker, reducing deployment cycle time and improving resource efficiency; contributed to code reviews and documentation following established software engineering practices.
• AI Service Performance Optimisation: Reduced average AI service response time by 20% through strategic caching layers and streamlined API logic; collaborated in Agile sprints with distributed ML and platform teams on banking and retail client solutions.
• RESTful API Design (OpenAPI): Designed and documented RESTful APIs following OpenAPI standards for AI model serving endpoints consumed by web and mobile frontends.

M.Tech - Information Technology - Indian Institute of Information Technology (IIIT), Allahabad (2023 - 2025)

B.E. - Information Technology - Prof. Ram Meghe Institute of Technology and Research, Amravati (2019 - 2022)