Nikhil Anand
I am an AI Researcher at Adobe Research, working on agentic evals and interpretability methods that make AI systems more reliable.
I graduated from IIT Madras with a dual degree (BTech in Bioengineering, MTech in Data Science), where my thesis, born out of an internship at Adobe, focused on improving context faithfulness in LLMs through their internal representations.
Outside research, I write about AI on Medium and on my Substack publication, AI Made Easy, turning hard ideas into intuitive visuals and stories. In my free time I read, play the guitar, sing, and play tennis.
News
- Jan 2026 [Paper] Our paper ContextFocus: Activation Steering for Contextual Faithfulness in LLMs is out on arXiv.
- Jan 2026 [Project] Launched PaperVideo: turn any research paper into a narrated animated walkthrough.
- Jun 2025 [Role] Joined Adobe Research full-time as an AI Researcher in Bengaluru.
- May 2025 [Degree] Graduated from IIT Madras with a Dual Degree (BTech + MTech), second in my class.
- Sep 2024 [Patent] Filed a patent at the USPTO based on internship work on contextual faithfulness in LLMs.
- Jun 2024 [Internship] Started a research internship at Adobe Research.
- Dec 2023 [Award] Bronze medal at Inter IIT Tech Meet 12.0 (Adobe’s Behaviour and Content Simulation track).
- Aug 2023 [Fellowship] Completed the MITACS Globalink research fellowship at the University of Toronto / Sunnybrook Health Sciences Centre.
Publications
ContextFocus: Activation Steering for Contextual Faithfulness in Large Language Models
arXiv preprint, 2026
A lightweight activation-steering method that improves how LLMs follow externally provided context when it conflicts with parametric knowledge. Minimal inference-time overhead, ~95% cheaper than finetuning while remaining interpretable. Evaluated on Llama and Mistral; improves faithfulness by up to 24% over ContextDPO and COIECD on the ConFiQA benchmark.
Experience
AI Researcher, Adobe Research
Working on evals for AI agents. Focused on making model behaviour more transparent, controllable, and reliable in production settings.
Research Intern, Adobe Research
Improved LLM faithfulness to provided context by up to 24% on Llama, Mistral, and Gemma using activation steering. Filed a patent at the USPTO and received a Pre-Placement Offer.
MITACS Globalink Research Intern, Lin Brain Lab, Sunnybrook Health Sciences Centre / University of Toronto
Built predictive models for brain age from EEG, fMRI, and structural MRI features (a biomarker correlated with several neurological diseases).
Research Intern, National Centre for Biological Sciences (NCBS)
Replaced a Monte-Carlo sampling pipeline for protein-complex structure prediction with a reinforcement-learning-based approach, with applications to identifying potential drug interactions.
Software Developer, Desklamp (Y Combinator W23)
Built the mobile PWA from the ground up using PDF.js and React — immersive home screen, custom PDF viewer with zoom/find, and an inline formattable text notebook.
Projects
A tool that turns any research paper into a short, narrated animated walkthrough, built to make paper comprehension faster. Converts arXiv PDFs into a video script, generates the animations frame-by-frame, and stitches them together with narration.
Behaviour and Content Simulation (code)
Predicted a tweet’s like-count from its content, author, and date, and generated tweet content given the like-count, date, and author. Our team secured the Bronze Medal among 23 IITs.
Brain Age Prediction from Neuroimaging
Built multi-modal predictive models combining EEG, fMRI, and structural-MRI features to estimate biological brain age, a biomarker tied to neurological disease progression.
Writing
I write about AI, including areas like mechanistic interpretability, deep learning, and RAG, on Medium and on my Substack publication AI Made Easy. A few pieces I’m proud of:
- A friendly guide to linear algebra for AI. The linear-algebra primer I wish I’d had in school.
- A comprehensive guide to AI. A single piece that organises all my work in one place.
- Anthropic’s circuit tracing technique. Sparsity, transcoders, and the interpretability work behind tracing Claude’s thoughts.
Extracurriculars
Head, Electronics Club, Centre For Innovation (CFI), IIT Madras
Co-led a managerial team of 10 coordinators and ~40 undergraduates running 10+ multidisciplinary projects across IoT, PCB design, and ML; plus campus-wide sessions and events. (Coordinator the year prior, May 2021 – Apr 2022.)
Music Contingent (Vocals), IIT Madras
Sang with the institute music contingent through my final year; it expanded my worldview of what music was in a way that’s hard to describe. Some of it lives on Instagram.
Education
Indian Institute of Technology, Madras
Dual Degree (BTech in Biological Engineering + MTech in Data Science). CGPA 9.50 (2nd in class). Music Contingent (Vocals), Institute Tennis Team.