I build AI systems end to end: agents, retrieval, ranking models and the infrastructure under them. Then I measure them against held-out data until they break, and I publish what didn’t work alongside what did.
That habit shows up everywhere: preregistered predictions in my quantum-ML research, a negative result published for my 3D-reconstruction platform, and upstream bug fixes in PyTorch, pandas and mem0 that started as “this number looks wrong.”
NowLeading the HackerRank Orchestrate October leaderboard, revising my quantum-ML paper for Springer, and looking for applied AI and ML roles.
#1of 50,000+HackerRank Orchestrate, Oct 2026live leaderboard
#38of 32,000+ teamsAmazon ML Challenge 2026team lead
Preprint · public on GitHubSubmitted to Quantum Machine Intelligence (Springer), Aug 2026Major revision
Geometry Conservation in Variational Quantum Classifiers and Where Neural Collapse Can Appear
G Karthik Koundinya (first author, GCET), Puja S Prasad (corresponding) and Partha Pratim Roy, IIT (ISM) Dhanbad
Prior work proved that variational quantum classifiers reach Neural Collapse only at the static optimum. I asked whether collapse actually emerges during training, and found that it happens by an exact rotation, not a contraction: training conjugates every state by one global unitary.
The rotation result holds to machine precision (~10⁻¹⁵ relative change) at 4, 6 and 8 qubits, on MNIST and Fashion-MNIST under two encodings, across a second optimizer and ansatz family. It was also verified on two IBM quantum backends. A matched classical MLP does the opposite: within-class scatter grows 27–163×.
The study ran as preregistered science: 52 predictions across 10 stages, every threshold committed before its metric was computed, on a frozen spec with 173+ tests.
From landing pages to production ML. Most recently an applied AI internship building autonomous research review.
Aug – Oct 2026
RamanIQ · Applied AI Engineer Intern · AI startup
Built REFEREE, an autonomous research-review product combining multi-agent orchestration and deterministic pipelines for paper analysis, experiment execution and evidence-backed verdicts; achieved ~87% accuracy on evals and tests.
+ 2 more
Screened 36K+ research papers across top venues (NeurIPS, ICML, ICLR, CVPR, ACL) and curated 380+ compute-feasible reproduction candidates by evaluating code and dataset availability, benchmarking selected findings against Hugging Face reproduction reports.
Contributed 19+ pull requests across Polymath and Orbicular, delivering data connectors, UI improvements and backend features in collaboration with the engineering team.
Built the preprocessing and feature pipeline behind a production REST API over 175K+ auction notices: CatBoost with GPU embeddings, 84.3% ROC-AUC, 80% less manual review.
+ 3 more
Shipped features to production serving 5K+ users per day.
Trained a CatBoost Cox survival model (0.79 C-index) and deployed it behind a validated Flask inference API.
Architected the initial hybrid RAG retrieval platform; co-built an AI scraping system that cut manual data collection ~80%.
Certifications · Microsoft AI-103 · Azure AI Solutions · Microsoft DP-800 · SQL AI Developer Associate · MongoDB Associate Developer · OCI Generative AI Professional · GitHub Foundations · Stanford / DeepLearning.AI ML Specialization