I'm 23. I grew up in India, came to Germany for the engineering, and stayed because the work turned out to be the interesting kind. Munich is home now, Heilbronn is often, and there's a boarding pass in the calendar more weeks than not. Hindi is my first language and English at C2 is the one I work in. My German sits between B2 and C1 and gets worse after nine. My Spanish is A2, which covers dinner and very little else.
The day job is agentic AI at appliedAI Initiative. The night job is an M.Sc. at TUM in natural language processing. In between I teach embedded systems to students who are younger than me and twice as fearless.
Most disagreements about AI are disagreements about what a thing would do once it exists. Those end fastest when somebody goes and builds the small version, so that's usually what I do first. It settles what a meeting can only circle, and it costs a great deal less than being wrong at scale.
I talk to people constantly, and mostly to people whose working day I don't understand yet. I sit in the TUM Investment Club and the Munich Finance Circle because knowing how a business makes its money is what separates a clever model from a useful one. Walking into an industry I don't know yet is the part of this I like most, which is lucky, because it's most of the job.
Newest first, career on one track and education on the other. Nearly every entry sits in the same place: where a model meets a company that already has a way of doing things.
Some of this shipped inside a company. Some of it I built because I wanted to see whether it would work. Some was built over a weekend against a clock. Ask about any of it and you'll get the version with the parts that didn't work still in it.
This isn't a publication list. It's the university work sitting behind the rest of it, graded, each piece ending in a number. All of it still comes up in client rooms.
Can a small model hold a 2D map in its head well enough to navigate it? I compared several on accuracy, hallucination rate and latency. That's the same three-way trade-off every company runs into the moment it asks why it can't just use the biggest model available.
The question was whether a large model can do more than notice a contradiction. Can it choose the right repair? Finding the conflict turned out to be the easy half, and that's why I never let an agent write into a system of record without a person in the loop.
Large models sound fluent in Hindi. The paper tested whether the fluency reaches the grammar, by making them tag parts of speech in Hindi and in English and comparing the gap. It's the clearest argument I have for testing a model in the language your business actually runs in.
Can a BERT-style model learn grammar faster if you make it tag parts of speech while it pretrains? I added a POS objective alongside MLM and NSP, then measured on BLiMP and GLUE in the Strict-Small track of the BabyLM Challenge. The constraint is the interesting part: a fixed and deliberately small data budget, which is the position most companies are in when they stop daydreaming about training their own model.
Some of this explains how I work and some of it just explains me. It's all here for the same reason: you'll spend real time in a room with whoever you hire, and it's fair to know who that is.
I dive whenever I can get to the sea. Everything down there moves slowly, me included, and I come back up in a better mood than I went down.
A pool will do, a lake is better. It's the one form of exercise I've never had to talk myself into.
The simplest thing I do. Nothing to book, nobody to meet, no equipment to forget. Most problems are a size smaller by the time I get back.
StravaI started as an adult, which is late, and I don't mind. My hands are still learning to cooperate.
I'm working towards a private pilot licence. Aviation got hold of me early and never really let go.
I grew up in India and I'm at home in Germany now. Both are true at the same time. The chilli in my cooking is the part that never adapted.
I like good clothes, and getting it right is a small pleasure I've never wanted to grow out of.
Markets are a hobby. A good evening is one where somebody changes my mind, and that usually happens at the TUM Investment Club in Heilbronn or the Munich Finance Circle.
I'm not casual about coffee, and I'll walk a bit further for a good cup.
I like talking to people, especially about work I know nothing about. Tell me about your job and I'll still be asking questions when the food arrives.
That's the whole record. Alongside appliedAI Initiative I take on a small number of transformation projects and workshops each year, which is what keeps me close to the work instead of talking about it. If any of it fits what you're doing, the next page is one email long.