M.S. Information Science · JAIST (Japan Advanced Institute of Science and Technology); Ishikawa, Japan
B.Tech Mechanical Engineering · SRM Institute of Science and Technology; Tamil Nadu, India
I design AI-powered systems under real-world constraints — edge hardware, zero cloud dependency, budget-first architecture.
- XAI-SwissKnife:
An intuitive tookkit that auto-routes the correct explanation method to the right model and target hardware. - Local multi-agent system using knowledge distillation:
Claude → DeepSeek R1 → Qwen 3.5-9B, Q4_K_M quantized; GTX 1060, 32GB RAM target
- Shipped an on-device inventory app with pruned MobileNet V4 Small
From model selection to APK in 3 days. - Built auto-quant, a benchmarking tool that helps quantize models based on type and level
bringing quantization techniques into one web based UI and providing comparison for different
levels along with Pareto Frontier graph to aid with best quantization level selection.
Architecture formulation to Proof of Concept deployment in 4 days.
Comparison of Different Interpretability Methods with Professional Annotated Data for ViT Based Medical Image Classifier: \
- XAI research for medical imaging
- Finetuned ViT, 90% diagnostic accuracy
- Performed benchmarking for six different XAI methods
- Cross validated against interpretation data from six radiologists
- Thesis available in JAIST reposiory
- Frame the problem with constraints (hardware, budget, timeline)
- Define success metrics upfront (latency <50ms, offline-first, privacy)
- Architect modular solutions with clear tradeoff documentation
- Validate with lightweight user testing or expert feedback
- Iterate based on Pareto-optimal decisions, not perfection