Berlin · Co-founder and CTO, Bryo · LLM experiments

Nikhil Mudholkar

I took LLM agents from prototype to production with industrial customers, and I run independent experiments on how LLMs and decision models behave in real work.

Now
Co-founder and CTO, Bryo.ai. AI quoting and order automation for German manufacturers. Raised about $4M, built a team of 10, and took the product into production with customers in Germany and the UK.
Research
Ontologies, solvers and AI for the industrial base. I led the design of an ECLASS-aligned product ontology across 1,000+ categories and of agents that drive product configurators, and I publish independent LLM benchmarks.
Speaking
Speaker, Odoo Experience 2023, Brussels. I have also spoken at events about AI and entrepreneurship, and I run workshops for manufacturers on how to use AI in their business.
Before
Data engineer, Capital One and Tier. In 2018, before LLMs, I wrote a language parser and transpiler with ANTLR4 that translated 11,000+ SAS scripts to PySpark.
Studied
IIT Bombay. B.Tech in Materials Science, minor in Computer Science.
2 letters of recommendation

Independent experiments · 2026

Experiments

Cloudflare's new extraction model reads a document and returns typed fields in one pass, no token-by-token writing. I tested the two-day-old preview on 503 order emails and 711 receipt photos against five LLMs. It is the fastest and most consistent, but it invents values that are not there and works out a receipt's covered total only a third of the time.

Oct 2026 · 503 emails · 711 receipts · 6 models · Read →

A backdoor hidden in a model's weights, invisible until it hears ten words. I fine-tuned one into an open LLM overnight for a few hundred dollars, then measured how reliable, how undetectable, and how shippable it is.

Oct 2026 · security research · Read → · Thread on X ↗

A decision model against two Gemini models. Jev is a little less accurate, costs a tenth as much, and is the only one whose confidence you can trust.

Sep 2026 · 1,565 emails · 3 models · Read → · Thread on X ↗

Ask the what-ifs before you choose. An 8-step agent loop drops to 3 calls and gets more accurate on both Jev and Gemini.

Sep 2026 · 300 requests · 8 arms · Read → · Thread on X ↗

Jev, Perplexity and Cloudflare Clef side by side. Perplexity wins overall. Clef is accurate but underconfident and slower than claimed.

Oct 2026 · 1,565 emails · 4 models · Read → · Thread on X ↗

Perplexity's new v1.1 and OpenAI's Decisions API join Jev and Cloudflare Clef on the same emails. Perplexity v1.1 is the most accurate and the cheapest. Jev is still the one whose "certain" you can trust.

Oct 2026 · 1,565 emails · 6 models · Read → · Thread on X ↗

Work and education

Experience

  1. Feb 2023 to now
    Bryo.ai · Co-founder and CTOBerlin, Germany
  2. Oct 2021 to Sep 2022
    Tier Mobility · Data EngineerBerlin, Germany
  3. Jun 2019 to Sep 2021
    Capital One · Data EngineerBengaluru, India
  4. Jul 2015 to May 2019
    IIT Bombay · B.Tech, Materials Science, minor in CS EducationMumbai, India
    1. Summer 2018
      Capital One · Intern, SAS-to-PySpark transpilerBengaluru, India
    2. Summer 2017
      Lokacart · Intern on a Microsoft and NABARD project, an e-commerce app for farmersMumbai, India
Download resume (PDF)

Skills

AI
Agent harnesses, evals and LLM-as-judge, search and retrieval, Vertex AI, LiteLLM, Langfuse
AI coding
Heavy Claude Code user, 5 to 6 hours a day across parallel sessions, plus Codex
Engineering
Python, SQL, TypeScript, Java · GCP, AWS, Terraform, Airflow, dbt, Postgres
Languages
English (fluent), German (B1)