Kanak Raj
Researcher and engineer building practical AI systems — from legal reasoning and knowledge graphs to agentic workflows and LLM alignment. Currently at Thomson Reuters Labs.
Completed a 6-week research residency at LossFunk, working on research problems at the intersection of deep learning and creative AI.
Building legal AI reasoning systems at Thomson Reuters Labs. Work spans synthetic data generation for Process Reward Models, IRAC knowledge graph construction from the Westlaw corpus, and end-to-end LLM workflows for legal document processing. Also built several internal tools including a real-time Truth Social monitoring system for Reuters journalists.
Worked with the AI4Code group on Programming with Representations (PwR) — building an intermediate representation layer between natural language and LLMs. Developed the NL2DSL translation pipeline using GPT-4, achieving an 85% reduction in hallucinations. Also contributed to Jugalbandi, an open-source chatbot platform featured in Satya Nadella’s keynote and later adopted by Bhashini across government initiatives.
Conducted research on multimodal knowledge graphs and recipe understanding with Dr. Amit Sheth and Revathy Venkataramanan. Defended my Master’s thesis on Knowledge Enabled Multimodal Ingredient Substitution — a knowledge graph with 27K ingredients and 40K substitution pairs, used in the UC Irvine + Stanford Health Hackathon 2024. Work also led to Cook-Gen (IEEE SMC 2023) and a submission to AAAI-25.
Visiting researcher in the Societal Computing group, working on satellite image super-resolution using temporal and multispectral data. Applied GAN and diffusion-based approaches for wildlife tracking and disaster analysis with mentors Ingmar Weber and Ferda Ofli.
Developed personalized response generation models under Prof. Manas Gaur as part of the HealthCareNLP grant. Used reward scaling over BART and T5 for knowledge and persona-aware generation. Work accepted as K-PERM at AAAI Spring Symposium 2024.
Built model optimization pipelines for accelerated inference using TensorRT and OpenVINO, enabling deployment of computer vision models on edge devices.
Implemented StyleGAN-based architectures for disentangled video interpretation across domains at the Video Analytics Lab under Rishubh Parihar.
Reached the national finals of Smart India Hackathon under ISRO guidance, building a real-time Cyclone Intensity Estimator using deep learning on INSAT-3DR satellite data.
Researched medical image processing at the Visual Learning and Intelligence Lab. Developed a novel architecture for dental X-ray segmentation, published as SFFNet at IEEE APSCON 2023.