Pranav Shinde
Computer Vision Engineer | ML Systems & MLOps
Badawe Engineers (Defence R&D)
AI Engineer - Computer Vision & ML Systems
- Built and maintained 500k+ image annotation pipeline with custom SAM 2.1 boundary refiner for automated mask correction.
- Trained 13-class YOLO instance segmentation model (mAP50~0.74) with WeightedDetectionLoss for severe dataset imbalance.
- Engineered dual-GPU training pipeline (RTX 5080/3060) resolving CUDA conflicts, reducing validation turnaround by 18%.
- Migrated inference PyTorch → ONNX/TensorRT with INT8 quantization — 9.2 GB → 2 GB memory (−78%), ~23 ms → ~3.7 ms latency. Profiled with Nvidia Nsight Systems/Compute.
- Implemented P2 feature pyramid heads to improve separability between visually similar classes in cluttered scenes.
- Automated full annotation workflow via CVAT REST API, replacing hours of daily manual effort.
Vital Vistara
Full Stack Intern (ML Backend)
- Designed AWS infrastructure for ML workloads: VPC, RDS, Lambda for serverless inference, S3 for artifacts.
- Implemented CI/CD with Docker builds, ECR/ECS Fargate deployment, and automated ML model versioning.
- Deployed scalable, cloud-native applications on AWS, achieving 99.9% availability.
Oneqid Technologies, IIT Delhi Research and Innovation Park
Web Development Intern
- Migrated legacy infrastructure to Next.js, resulting in a 35% increase in page load performance.
- Implemented automated Vercel deployment workflows, boosting Lighthouse performance scores from 65 to 92.
2026
Frontend Engineer | Engineered a high-performance interactive node graph workspace using React Flow 12 for dynamic node rendering and connection edges. Managed complex UI state and cache synchronization using Zustand and TanStack React Query, backed by a mock service layer (MSW) for API simulation. Designed a pixel-perfect, theme-reactive UI with interactive resource sliders and custom color-coded status badges utilizing Tailwind CSS.
Sep 2024 - 12/2024
Full Stack Engineer | Engineered secure sandboxed code execution using Docker with CPU/memory limits on AWS ECS Fargate. Implemented PostgreSQL connection pooling (PgBouncer), reducing connection overhead by 70%. Built scalable backend for supporting 1000+ concurrent users with strict security isolation.
Aug 2024 - Oct 2024 • Top 10 National Finalist, Smart India Hackathon 2024
Backend & ML Engineer | Deployed on-device pose estimation (Google ML Kit) on Android for low-latency, privacy-preserving fitness evaluation. Engineered Firebase-backed pipeline for nationwide athlete profiling and leaderboards supporting 50,000+ competing teams nationally.
2026
Built a 4-stage automated pipeline to survey 500k+ datasets on Hugging Face Hub — metadata harvesting, fuzzy deduplication, VLM-based visual classification (Qwen2.5-VL) with majority voting, and zero-download remote Parquet schema inspection. Screened 500k+ items in under 5 hours.
2026
CLI tool that benchmarks your YOLO model across every export format available on your machine — PyTorch, ONNX, CoreML, TensorRT — and produces real FPS, latency, and accuracy-delta numbers. Motivated by daily work profiling INT8/TensorRT pipelines. Waitlist open.
2026 · Open source contribution
Diagnosed macOS kernel panics (IOGPUMemory.cpp) when serving a 27B model at 64k context on 24 GB unified memory, tracing the root cause to step=256 KV cache reallocation churn. Patched mlx-lm to preallocate the full KV buffer in a single mx.zeros call — zero reallocations, zero wired-memory spikes, stable 65,536-token serving. PR #1920 open against ml-explore/mlx-lm.