Vision systems that look at Earth from orbit and return verified, measurable answers — not just pixels.
Currently engineering at Stellaria.
| Stellaria IOD | Interactive object detection for satellite & aerial imagery. Click any object → it is identified and measured. Type a word → every instance is found across the scene at native resolution. Open a second scene → the same search runs side‑by‑side.SAM 3 CLIP DINOv3 GeoTIFF / COG |
| Murph | A satellite-imagery agent: plain-language query in, verified objects out — masks, real measurements (length · width · wingspan · heading · area), per-object attributes, PDF reports, and mask-true GeoJSON in WGS-84.Agents × SAM 3 FastAPI GPU sidecars |
| geochange | Object-level temporal change detection. Two passes over the same location → a georeferenced, classified change set: still there, moved, grew, shrank, or replaced — every verdict with evidence and confidence attached, not a raw difference mask.Bi-temporal CV DINOv3 Change heatmaps + PDF reporting |
| Foundation-model R&D | Training-free in-context segmentation on frozen DINOv3, LoRA fine-tuning for domain adaptation, and embedding-space retrieval designed to survive backbone swaps.DINOv3 LoRA PyTorch |
| voice-rag | Multilingual (en/hi) voice-driven RAG over MS MARCO-XI — sentence-aware chunking, FAISS HNSW, cross-encoder reranking. Retrieval core measured at P50 ≈ 28 ms, honestly benchmarked end to end.FAISS e5 embeddings Ollama |
Also recently shipped: Knack — an AI-scored, gamified soft-skill trainer for engineering students.
|
Staged Embarrassment Learning (v2) 99.1% FLOPs reduction for ResNet training — bio-inspired, 100% local  
|
Recylo AI Waste Segregator Patented Edge AI on Raspberry Pi 5 — 90% classification accuracy at source 
|
IWM Reverse Vending Machine Automated PET recycling — JVTM prize winner, ₹1,00,000+ funded 
|
OceanOS IoT Simulation Marine waste management — Three.js simulation, dashboard & IoT app 
|
I build vision systems that understand the physical world — from satellite scenes measured object by object, down to models squeezed onto a Raspberry Pi. Information Science & Engineering student at BGSCET, Bengaluru; currently turning raw orbital imagery into verified, measurable answers at Stellaria. I care about architectural efficiency, honest benchmarks, and shipping on real hardware.
| Patent granted | AI-driven waste segregation system — Design No. 482604-001 |
| Funding secured | ₹1,00,000+ for sustainable AI development |
| SIH 2025 Grand Finalist | 120-hour hardware + software edition |
| Awards | JVTM National Exhibition Winner · ZeroOne Runner-Up |