Practitioners, not presenters

Khushi is completing a B.E. in Artificial Intelligence & Data Science at Zeal College of Engineering and Research. Her project work covers computer vision — a gesture-controlled virtual mouse built with Python and OpenCV, using real-time hand landmark detection — and full-stack web development, including a multi-vendor service aggregator with authentication, payments and an admin panel.

She served as Vice President of the SESA committee in her department and took part in the Smart India Hackathon.

At Deesha she mentors learners on the AI and machine learning path: framing a first problem, getting a model working end to end, and the gap between a notebook that runs and a project someone else can run.

Suhani is reading for a B.Tech in Chemical Engineering at the Indian Institute of Technology Jodhpur, ranked 4th in her department, alongside research and project work in computer science and machine learning.

As an undergraduate researcher in the Department of Computer Science, supervised by Prof. Tanmay Inamdar, she implemented and benchmarked four algorithms for the NP-hard vertex cover problem — GreedyMIS, LP rounding, NeighborCover and a parallelised variant — across 588 runs on Erdős–Rényi graphs of up to 5,000 vertices. Over 49 size-and-density configurations, GreedyMIS ran fastest in 94% of them while LP rounding produced the smallest covers. That distinction between the fast answer and the good one is exactly the judgment a learner needs and rarely gets taught.

For Inter-IIT Tech Meet 14.0 her team built a conversational analytics platform over 19,621 call-centre transcripts and more than 686,000 messages, orchestrating a 13-node LangGraph agent with query decomposition, multi-hop reasoning, and routing between text-to-SQL and hybrid retrieval combining BM25, dense search and reranking. It beat every RAG baseline on a 100-query benchmark — 8.85 out of 10 for relevancy against 4.63, and 9.21 for citation fidelity — human-validated at a Pearson correlation of 0.76. The team placed 8th among 23 IITs.

Her other work spans deep learning and classical computer science. She trained a VelocityGAN-style CNN encoder–decoder in PyTorch to reconstruct subsurface velocity maps from seismic waveforms, reaching SSIM 0.71 on a 6,000-sample holdout by balancing L1 against adversarial loss. She built a C++ autocomplete engine on a radix trie, BK-tree and min-heap that returns ranked suggestions in under 0.25 ms per keypress, and cut repeated-query latency twelvefold with a constant-time session cache. An earlier course project implemented Dijkstra's algorithm and k-means from scratch over a Haversine-distance graph.

She also architected the INSPECT vehicle damage assessment pipeline — ResNet-50 classification, YOLOv8 detection and Qwen2.5-VL reporting over a 15,000-image corpus — and directed a five-member team through model selection, evaluation and deployment-oriented design across all three stages. She is an active member of RAID, the AI society at IIT Jodhpur.

At Deesha she mentors learners on algorithms, machine learning and computer vision — close enough to the work they are attempting to remember exactly where it goes wrong.