Currently Building
Practical LLM and agentic systems for legal workflows, with strong reliability and validation loops.
Kanak Raj
AI Engineer and Research Enthusiast
Building reliable machine learning systems and practical AI products.
I work across deep learning, product engineering, and open source. This site is where I share technical write-ups, project updates, and learning notes from real-world ML and software work.
Research Interests
My interests span computer science and machine learning, especially deep learning for high-impact domains and robust production systems.
- Computer Vision
- Natural Language Processing
- Generative Models
- LLM Reasoning
- Harness Engineering
- Memory & Adaptive Agents
Quick Links
Personal Touch
What I Care About
Systems that are useful in production: grounded outputs, measurable quality, and human-in-the-loop trust.
Beyond Work
I enjoy writing technical explainers, open-source experiments, and making complex topics easy to follow.
Recent posts
The Disappearing Parliament: What India’s Legislature Actually Does With Its Time
India’s parliament sat for 55 days a year in the 17th Lok Sabha — the lowest of any full-term parliament. 35% of bills passed with under an hour of Lok Sabha...
The Fragmented Calendar: A Data Atlas of Global Stock Exchange Holidays 2025
Japan loses more trading hours to holidays than any other major exchange. India doesn’t have the most holidays — but it has the most fragmented ones. The pat...
Change Detection: Extracting Ground Truth from Noisy Edits Permalink
You want to build an AI system that automatically updates documents when new information arrives — a court ruling updates a legal article, a product release ...
The Long Game: The Technical and Strategic Reality of India’s PFBR Criticality
Every PFBR headline just repeated ‘Three-Stage Nuclear Power Programme’ without explaining what the stages mean or what flows between them. This post is what...
Fine-Tuning LLMs with Non-Differentiable Human Feedback + RL
The blog is focused on giving you an intuition of why even using non-differentiable reward function we are able to use Group Relative Position Optimization (...