Building systems that see, reason, and decide —
from satellite orbits to emotional intelligence.
AI/ML engineer at the intersection of research depth and engineering rigor. Twice selected for ISRO problem statements at Inter IIT. 2000+ DSA problems solved. I don't just write models — I build pipelines that work under real constraints.
I'm a third-year undergraduate at IIT Bhilai, studying Data Science & AI — one of India's most rigorous DS/AI programs. My work sits at the overlap of computer vision, NLP, and systems engineering, with real deployments for ISRO and competitive hackathon environments.
I operate at every level of the stack: from raw NumPy pipelines over Chandrayaan-2 XRF data, to training multimodal transformers, to shipping REST APIs with FastAPI. I care about results, not just methods.
My competitive programming habit isn't separate from my AI work — it sharpens the way I think about complexity, edge cases, and system design. 2000+ problems solved isn't a number — it's a problem-solving posture.
Designed and built an end-to-end vision-language system capable of image captioning, Visual Question Answering, and text-driven object grounding — all on satellite data. The core challenge: rotated objects at arbitrary angles and extreme scale variability across orbital imagery. Addressed this with a hybrid pipeline coupling YOLO-OBB (Oriented Bounding Box) detection with a multimodal fusion architecture, served via FastAPI. Evaluated rigorously on IoU and BLEU-3 against GPU benchmarks under competition constraints.
Processed raw CLASS XRF spectrometer data from Chandrayaan-2 to generate high-resolution elemental composition maps of the lunar surface. Designed a custom clustering and spatial filtering pipeline to denoise orbital measurements, then leveraged PyTorch parallelization to cut processing time by 40%. Integrated QGIS and Astropy for geospatial analysis and astrometric calibration. Delivered production-quality lunar maps under competition deadlines.
Built a dialogue model that actually understands emotional context — not through keyword matching, but through COMET-based commonsense reasoning integrated with persona-conditioned generation. Introduced the Persona Attention Loop (PAL) architecture to maintain consistent persona across conversation turns. Trained in two stages: MLE on a pointer-generator network, then REINFORCE-based RL to optimize for non-differentiable dialogue quality metrics. Outperforms baselines across BLEU-4 and ROUGE-L.
Designed and shipped a full productivity platform during The Forge hackathon (Meraz 6.0). Core innovation: a custom Focus Efficiency Algorithm that quantifies user behavior patterns — beyond simple time tracking — to surface meaningful productivity insights. Built on React + Vite, deployed and demonstrated live. Secured 3rd position.
2000+ problems across Codeforces, CodeChef, LeetCode, and GeeksforGeeks. This isn't a metric — it's a mindset. Every problem is a lesson in edge-case awareness, time complexity intuition, and clean implementation under pressure. The same habits that make competitive programmers sharp make engineers reliable.
Open to research collaborations, internship opportunities, and high-impact roles in AI/ML engineering. If you're working on problems worth solving, I want to hear about them.
Open to internships and high-impact roles in AI/ML and software engineering, where I can work on real-world systems and challenging problems.