I'm a researcher and builder from Hyderabad. I like finding out what models really do, and making them fit on hardware people can afford.
Now: Research Intern on WISER, an Indo-German drone project for spotting pests in paddy. My part is multimodal data and sensor fusion.
Also: checking whether a sepsis model's explanations predict what it does, with Prof. Hamdi Al-Jamimi at KFUPM.
Building: Vak, AI literacy tools in Telugu and Hindi.
Open to: working together on interpretability, edge ML and model verification.
- Interpretability: pulling the small circuit behind one behaviour out of a language model. HAGD ships with a public audit of what its code reproduces.
- Edge ML: ResNet-18 pruned from 11.2M to 2.5M parameters, 142 ms per image on a Raspberry Pi 4. PMP-DACIS
- Efficient LLM agents: each step runs at FP16, INT8, INT4 or INT2. The paper reports 68.3% less compute. AgentCompress
- Checking claims: four detector upgrades that each helped alone but hurt together. LAF-YOLOv10
- Tools: LaTeX Intelligence turns LaTeX into an editable Word file. Guardian-AI filters prompt injections out of security logs.
