Most AI projects never leave a notebook. I build ones that ship.
I'm Atharva Sarde β a sophomore AI/ML engineer who bridges the gap between a model that works in Colab and a system that works in production. Currently interning at RAMS Inspect, taking YOLO-based defect detection all the way from raw dataset β quantized INT8 β real-time inference on edge hardware.
I don't just train models β I deploy them. I don't just write code β I ship products.
current_role = {
"company": "RAMS Inspect",
"title": "Artificial Intelligence Intern",
"since": "April 2026",
"stack": ["YOLO", "PyTorch", "ONNX", "MLIR", "INT8 Quantization", "MaixCAM / CVITEK NPU"],
"mission": "End-to-end CV pipeline: dataset β trained model β edge-deployed .cvimodel"
}Right now I'm:
- π― Training YOLO for real-time defect detection on industrial hardware
- β‘ Converting models:
PyTorch (FP32) β ONNX β MLIR β INT8 .cvimodel - π§ Deploying on MaixCAM / CVITEK NPU β actual edge, actual latency, actual results
- π Tuning mAP via loss function optimization and class rebalancing (under + oversampling)
Because a wrong diagnosis isn't just a bug β it's a life.
ResNet50 Diagnostician β MLP Auditor β Circuit Breaker Safety Protocol
β β β
84.84% Accuracy Entropy + Embedding Zero Silent Failures
(BreaKHis Dataset) Distance Introspection in Diagnostic Decisions
- Engineered a dual-stage deep learning framework β diagnostician + auditor
- Pioneered an AI introspection mechanism that captures 75% of model misdiagnoses before they reach a doctor
- Deployed Grad-CAM explainability β real-time heatmaps showing exactly where the model looks
- Designed a Circuit Breaker safety protocol that decouples prediction from trustworthiness
π Why it matters: This isn't a demo. It's a reliability-first medical AI system built for production safety.
Aligned with UN SDGs. Built for the people who need it most.
- Full digital health record system with multi-role portals + Row-Level Security
- AI-driven patient surge prediction improving hospital capacity planning by 35%
- Offline-first PWA for low-connectivity regions with real-time sync
- NDHM/ABHA compliant β national health data interoperability, built-in
GPT-4 powered. JWT secured. 99.5% uptime.
- Real-time code suggestions via OpenAI GPT-4 API
- Secure JWT auth + Supabase PostgreSQL + role-based access (users, mentors, admins)
- Sub-200ms latency | Responsive chat UI | Optimized deployment
| π Credential | π’ Issuer |
|---|---|
| AWS Cloud Certifications | Amazon Web Services |
| Introduction to Node.js (LFW111) | Linux Foundation |
| Data Analytics Job Simulation | Deloitte (Forage) |
| Bring AI to Work Workshop | |
| Certificate of Appreciation β Creative Lead | Universal AI University |
| Mumbai Hacks 2025 Participant | National AI/ML Hackathon |
β Most students: Train model β Accuracy looks good β Call it done
β
Atharva: Train model β Safety audit β Edge deploy β Monitor in production
| Dimension | My Approach |
|---|---|
| AI Safety | Built introspection layers that catch failures before they cause harm |
| Edge AI | INT8 quantization, ONNX/MLIR pipelines β not just cloud inference |
| Full-Stack | End-to-end ownership from ML model to deployed UI |
| Domain Depth | Medical AI, Industrial CV, Agentic Systems β not just tutorials |
| Production Mindset | 99.5% uptime, sub-200ms latency, NDHM compliance β shipped, not staged |
π¬ "I'm actively seeking internship roles in AI/ML Engineering and Software Engineering for 2025β26. If you're building something that matters and need someone who can go from idea to deployed system β let's talk."

