Harsh Upadhyay on Tal
Headline
AI/ML Engineer | Computer Vision | LLMs | Geospatial AI | Researcher
About
I am a B.E. CSE (AI & ML) student with hands-on experience in computer vision, deep learning, multimodal LLMs, RAG systems, and geospatial AI. I focus on building real, end-to-end AI systems that combine research depth with full-stack engineering. My major work includes: 🔹 Hallucination Hunter — an enhanced, training-free hallucination detection framework for multimodal LLMs • Built using BLIP-2, GroundingDINO & CLIP • Achieved +30.6% accuracy improvement on the POPE benchmark • Reduced inference latency by 4× using GPU parallelization, batching & caching 🔹 GeoLLM — a full-stack Geospatial AI system • Supports NDVI, LST, time-series analysis, flood mapping & LULC • Uses Google Earth Engine, RAG, FAISS, NER & intelligent intent routing • Built with FastAPI microservices + Next.js + MapLibre 🔹 Published Research (ICDSAI 2025, Scopus Indexed) • Improved YOLO using a ResNet50 backbone + rotated bounding boxes • Achieved significant gains on DOTA v1.5 I love solving AI problems that mix vision, language, geospatial intelligence, and real-world deployment. Seeking roles: AI Engineer, ML Engineer, LLM Engineer, Computer Vision Engineer, Research Intern, Applied ML.
AI stack
hugging face
langchain
ollama
openai
Work experience
AI Researcher — Hallucination Hunter at Independent Researcher, Jul 2025 - Jan 2026 · 7 mos, Mohali district, India, Built a training-free hallucination detection pipeline using BLIP-2, CLIP & GroundingDINO Achieved +30.6% accuracy on POPE benchmark Reduced inference latency by 4× with GPU batching Designed Hallucination Confidence Scoring (HCS)
AI Engineer — GeoLLM at Independent Product, May 2025 - Oct 2025 · 6 mos, Mohali district, India, Developed a full-stack Geospatial AI Assistant Integrated GEE + FAISS RAG + FastAPI microservices Built NER → polygon extraction → intent routing pipeline Created Next.js + MapLibre frontend with interactive tiles
Research Intern — YOLO-ResNet50 (Published Work) at Independent Researcher, Nov 2024 - Mar 2025 · 5 mos, Mohali district, India, Enhanced YOLOv5 using a ResNet50 backbone + rotated bounding boxes Achieved significant mAP gains on DOTA v1.5 Published in ICDAM 2025 (Scopus Indexed)
Education
B.E in CSE (AI & ML), CHANDIGARH UNIVERSITY
Intermediate, Rani Laxmi Bai Memorial School (R.L.B.)
Public links
github: https://github.com/HarshUpadhyay2003
