Janu Verma

I study how machines learn, remember, and reason. I strive to understand things deeply, and to share what I find. Over the last few years my focus has moved towards building learning systems that reason, adapt, and improve through feedback. Currently I am a principal applied scientist working on AI personalization at Microsoft, in London.

I trained in mathematics and physics, and a lot of my time still goes to them: reading, working through problems, and following science more generally. I have spent twelve years turning research into systems people use: in genomics, healthcare, payments, social media, and now agentic AICornell, IBM Research, Mastercard, Hike, Cult.fit, Microsoft. Cambridge and Kansas State before that. The longer version is below.. I write Incomplete Distillation, a research journal, and build small, complete systems to check what I think I know.

substack github scholar linkedin x email

Now

Working on: Personalization for Copilot by day. In public, the physics of coffee brewing, one method at a time: espresso and pour-over are written up, and the wet bed is still the open problem.
Reading: NW, Zadie Smith·The Pearl, John Steinbeck·Orbital, Samantha Harvey
Top of mind: The cosmological constant and the heat death of the universe·Type theory and the Curry–Howard correspondence·Temporal event sequences

as of September 2026

Projects all projects →

Bean Miles: coffee journeys drawn on a dark globe

Bean MilesSix months of coffee, measured. Vision LLMs read the bags; a globe replays the journeys.2026·07 · data viz · live

Stellar Evolution: every black hole begins as a breath of dust

Stellar EvolutionScroll, and ten million years pass. An explorable explanation of one star's life.2026·06 · explainer

LLMs in RecSysA six-part research arc from prompt-based preference prediction to a repo-backed cross-encoder.2026·06 · research

Marginalia: a conversational book recommender

MarginaliaDescribe a feeling, a question, a mood. A librarian who learns your taste answers with two or three books.2026·05 · tool

Also: Instagram Strategist, a content-strategy agent on Gemma 4. The projects group into three tracks: Personal Intelligence, RecSys, Explorable Explanations.

Writing januverma.substack.com →

2026·09·11Physics of Coffee Brewing, II: Pour-over MethodsEspresso has a pump. Pour-over has only a water level that rises and falls, so the whole brew is a transient. How to measure it with a flat-bottomed brewer and a stopwatch.

2026·08·07Physics of Coffee Brewing, IDarcy's law and the Kozeny–Carman equation get espresso's dry bed right. Real machines disagree, and the wet bed is why.

A single path drawn through a field of tangled dotted routes: accompany, not classify

2026·06·04Personal IntelligencePopulation systems ask what people like you enjoyed. Personal systems ask what you are becoming, and use that to decide what you should read next.

Incomplete Distillation is a research journal: long pieces, usually with code and a small experiment behind them, on whatever I am trying to understand that monthThe name is a confession. Each essay is a partial distillation of a subject that deserves more; a series continues until the residue is small enough.. Thirty-four essays since June 2025, mapped below by thread.

Every essay since June 2025, by thread; thirty-four so far. Hover for the title, click to read. The red mark is the newest.

Recommendation & personalization9 essays: Personal Intelligence2026·06

Marginalia: a conversational book recommender2026·05

RecSys after LLMs: four paradigms for what comes next2026·01

RecSys 2025 recap2025·10

Also: Joint retrieval and recommendation · Contextual bandits · Semantic IDs, a practical study · Semantic IDs, a deep dive · An LLM cross-encoder
Agents & reinforcement learning7 essays: I gave Gemma 4 my Instagram reels and built a content-strategy agent2026·04

From custom parsing to MCP: rebuilding the data-analysis agent2026·01

Building a data-analysis agent2026·01

LLM-based agents: towards agentic AI2025·12

Also: Multi-turn tool use with RL · Fine-tuning LLMs with RL, II: RLOO to GRPO · Fine-tuning LLMs with RL, I
Models & architectures9 essays: Research Briefings: Gemma 4, TurboQuant, Mamba-3, V-JEPA 2.12026·03–04

Diffusion models: I, from thermodynamics to language · II, inside Stable Diffusion · III, language models2026·02–03

Graph foundation models: building GPT for graphs2026·02

Graph Transformers: from message passing to global attention2025·07
AI × science5 essays: Physics of coffee brewing: I, espresso · II, pour-over2026·08–09

Protein design: a toy case study2025·09

From silver to gold: AI at the International Mathematical Olympiad2025·07

The protein folding problem: why shape is life's code2025·07
Essays4 pieces: How I became an amateur marine biologist2026·06

What a Time to Be Alive: the golden age of the curious mind2026·02

Claude built my website2026·02

On New Year's resolutions2026·01

Research google scholar →

Research has been my way into each field: crop genomics at Cornell, machine learning for healthcare at IBM Research, graph learning and semi-supervised methods for fraud at Mastercard, recommendation and personalization at Hike and Microsoft. Much of it shipped rather than published. Selected papersAlso patents: AI methods for predicting account-level risk of cardholders (Mastercard, filed 2022 and 2025), and mechanism-of-action derivation for adverse drug reaction prediction (IBM, 2019).:

Generative Recommenders for Zero-Query Prompt Recommendation. MLADS 2025 with V. Kolesnyk, R. Ronen
A Closer Look at Consistency Regularization for Semi-Supervised Learning. CODS-COMAD 2024 S. Ghosh, S. Kumar, A. Kumar, J. Verma
Guided Self-Training Based Semi-Supervised Learning for Fraud Detection. ICAIF 2022 A. Kumar, S. Ghosh, J. Verma
Label-aware Sampling using Contrastive Learning for GNN-based Fraud Detection. KDD 2022, Machine Learning in Finance J. Verma, G. Arora, A. Patankar, A. Chaudhry
Heterogeneous Edge Embedding for Friend Recommendation. ECIR 2019 J. Verma, S. Gupta, D. Mukherjee, T. Chakraborty
Cassava Haplotype Map Highlights Fixation of Deleterious Mutations during Clonal Propagation. Nature Genetics 2017 P. Ramu et al.

Speaking and advising invite me →

I speak about personalization and recommendation, agentic AI, and what separates robust AI systems from demos. Open to guest lectures and teachingAlso spoken at IIIT Delhi, University of Delhi, and the Institute of Mathematical Sciences, Chennai..

2026·05Preparing for an AI-First WorldKeynote · International Conference on Manoeuvring Business, Society and Culture, Jammu
Challenges in Retrieval for Enterprise AI AgentsGuest lecture · Applied AI, Rutgers University
LLMs in Recommender SystemsPyData London
Graph Embedding Methods for RecommendationData Hacks Summit, Analytics Vidhya
Transfer Learning in NLPPyData Delhi
Beyond QWERTY: Solving the Input Problem of IndiaFuture of Work, YourStory

I also advise founders on getting from an AI concept to a testable, dependable product: technical advisor to sia.vision, London, on agentic systems for storytelling and visual media. Open to advising UK founders working on agentic AI or personalization.

About

I started where the abstractions are purest: Cambridge, studying the mathematics of string theory and geometryMMath, Part III of the Mathematical Tripos, with a dissertation applying gauge/gravity duality to fluid dynamics. Then an MS in Mathematics at Kansas State, on geometric invariants arising from supersymmetric quantum field theories, and three years of a PhD before leaving to build things.. That search took me through computational biology at Cornell, drug discovery and clinical NLP at IBM Research, graph-based fraud and risk models at Mastercard, where I led a team of scientists and engineers, and personalised recommendation at Hike, Cult.fit, and now Microsoft.

The maths never left. It shaped how I think about representation, structure, and abstraction, and it shows up as a need to actually understand something before I trust it. The domain changes. The way of thinking does not. Writing is how I hold myself to that: putting an idea down in full, with the code beside it, is the fastest way I know to find out whether I understood it.

Off duty: pour-over videos, recipes, pop culture, and thinking too much about what to wear.