Creating the future Kumo infrastructure with Rust and using some really fun algorithms to do it.

Establishing a scalable explanation system for gnns too.

Publications:

1. Jain V, Alves Feitosa F, Kreiman G (2024). Is AI fun? HumorDB: a curated dataset and benchmark to investigate graphical humor. ICCV 2025 arXiv: 2406.13564.

2. Chaudhary I, Jain V, Singh G (2024). Decoding Intelligence: A Framework for Certifying Knowledge Comprehension in LLMs. AISTATS 2026, SeT LLM@ICLR 2024, arXiv: 2402.15929 [cs.AI].

3. I Chaudhary, V Jain, A Singh, K Sachdeva, S Ranu, G Singh (2025). Lumos: Let there be Language Model System Certification. arXiv preprint arXiv:2512.02966

**Focal Lab**

Implemented probabilistic certification framework for LLM knowledge comprehension via Knowledge Graphs using PyTorch/Huggingface, evaluating 8 models on few-shot medical QA via distributed inference on supercomputing clusters.

Identified adversarial input distributions that cause up to a 15% performance drop in models (e.g., Gemini-1.5-Pro) under natural prompt noise, informing robust model design.

**Ge Lab**

Improved Sampling of discrete diffusion based Graph Neural Networks (GNNs) models for inverse protein folding with information theory and entropy.

Scoped and built a deterministic end-to-end, full-stack code generation system using compiler principles to automatically translate Kumo UI entities into Python SDK code reducing manual authoring time by over 90%.

Designed and executed 120+ experiments and reduced search space by 75% for 10 hyperparameters for Graph Transformers (GT).

Developed synthetic datasets and trained custom GNN models outperforming baselines by an average of 10% for pre-sales engagements with about 5 enterprise clients.

Enabled large file upload feature supporting > 1GB local uploads, resolving critical pain points for multiple enterprise customers.

Delivered customer-facing tutorial notebooks for Kumo-RFM product release.

Developed CI/CD integrated comprehensive testing framework with synthetic test-data generation for LLM and RAG pipelines, improved performance by 20%.

Pioneered CLI tool and 4 API endpoints for batch transactions across 4 data asset types. Implemented automatic GraphQL query generation to accelerate development.

Worked with Azure OpenAI, TypeScript, Jest, CI/CD, Python, PyTest, MongoDB, GraphQL.

Introduced Novel Dataset on humor detection for Vision and Vision-Language models. arXiv: 2406.13564

Worked with PyTorch, DeepSpeed, Huggingface for model training and evaluation. For human data collection, designed web psychophysics experimets with JavaScript, AWS Lambda, HTML/CSS, and Amazon S3.