Applied AI Tech Lead · Munich

Kehan Liu

I build AI that ships.

Hands-on tech lead with 9+ years in AI/ML — fine-tuning transformers before ChatGPT existed, now shipping advanced RAG, agentic systems and custom LLM applications that enterprises run on.

Kehan Liu

9+

years in AI & ML

500+

FTEs of work automated

€50M+

annual business value

6

AI engineers led

Scroll

The Ascent

Every summit starts at base camp.

From energy engineering student to co-founder to leading enterprise AI — a decade of climbing, one camp at a time.

  1. 2012 – 2019Base camp

    B.Sc. Energy & Building Technology · M.Sc. Sustainable Energy Supply

    Beuth University Berlin / Tongji University Shanghai · RWTH Aachen University · Berlin, Shanghai & Aachen

    • B.Sc.: building-energy systems, simulation and HVAC — the engineering foundations.
    • M.Sc. at RWTH Aachen: machine learning, simulation, renewable energy and energy trading.
    • Two countries, three cities, one direction: energy.
  2. 2016 – 2019First climb

    Co-Founder / Data Scientist

    Joinergy · Aachen & Shanghai

    • Co-founded Joinergy alongside the master's: ML algorithms for electricity-load and PV forecasting for power retailers.
    • Designed energy-trading strategies for futures and spot markets.
    • Owned enterprise customers end to end as a co-founder — pre-sales scoping, technical sales, delivery.
  3. 2019 – 2023The traverse

    Data Scientist

    TenneT TSO GmbH · Munich

    • Production ML for photovoltaic and hydropower forecasting — 15-minute operational predictions for the grid.
    • Built TenneT's first MLOps pipeline; shipped with Python, PySpark, Docker, Jenkins and CI/CD.
    • Fine-tuned a DONUT document-understanding transformer in 2022 — months before ChatGPT.
  4. 2023 – todaySummit push

    Lead AI Engineer

    TenneT TSO GmbH · Munich

    • Hands-on tech lead and IC: writing, reviewing and optimizing the critical code on complex GenAI projects while leading a team of 6 AI engineers.
    • Tech Lead of the AI Coordination Office: setting company-wide AI strategy and the architectural standards for how the organization builds AI.
    • Shipping advanced RAG, agentic workflows and custom LLM applications on Azure — from self-hosted open-weight models to frontier LLMs.
    • Owning the customer relationship from pre-sales scoping through PoC to production adoption.

Selected Work

Systems that earn their keep.

Not demos. Production systems with users, uptime and a number attached.

TenneT · 2020 – present · Flagship

Intelligent Document Control

500+ FTEs saved · €50M+ annual value

Document-processing product that evolved across the modern AI era: CNN and SVM classifiers, a fine-tuned DONUT transformer (2022 — months before ChatGPT), open-weight Llama 3.1, and today frontier GPT-class models. Reviews thousands of technical documents weekly across 21 large grid projects.

  • Python
  • PyTorch
  • Hugging Face
  • DONUT
  • Llama 3.1
  • Azure
  • Kubernetes

TenneT · 2024 – present

Enterprise RAG Knowledge Assistant

Thousands of employees, one assistant

RAG systems that give employees a single chatbot over the organization's collective knowledge — querying Confluence, a lessons-learned platform and more, with grounded, sourced answers.

  • Advanced RAG
  • LLMs
  • Azure

TenneT · 2024 – present

Contract & Claims Agent

Accept or explain — always with rationale

An agent that evaluates submitted claims against contracts and legal documents via agentic orchestration, returning an accept-or-explain decision with a documented rationale.

  • Agentic Orchestration
  • LLMs
  • RAG
  • Azure

TenneT · Platform

GenAI & Agentic AI Platform

Data-sovereign AI for an entire company

Company-wide GenAI platform serving hosted LLMs — including self-hosted open-weight models — through an AI portal, API management, OpenWebUI and chatbot-as-a-service, so colleagues build agents and prototypes in a shared, data-sovereign self-service environment.

  • LLMs
  • Self-hosted Open-weight
  • OpenWebUI
  • RAG
  • API Management
  • Azure

Independent · Open Source

Tiny GPT in 96 Lines

A whole GPT, readable in one sitting

Karpathy's minimal GPT compressed to 96 lines of pure Python — training and inference intact, built for teaching how LLMs actually work.

  • Python
  • Transformers
  • Education
Keynote: “The AI Wave — how we face the change”

Keynote: “The AI Wave — how we face the change”

On Stage

AI is a wave. I help people surf it.

I give talks on GenAI, agentic systems and how organizations adapt to the AI wave — in English, German and Mandarin. From enterprise workshops to keynotes, I translate the frontier into something teams can act on.

Off the Clock

I help people go up — and down. Up, as in productivity with AI. Down, as in the slopes: when I'm not shipping AI, I'm in the Alps as a licensed ski instructor. Teaching someone their first turn and coaching a team through their first LLM deployment have more in common than you'd think: trust, small steps, and knowing when to let go.

First lift, fresh corduroy, somewhere above the clouds.

Toolbox

Depth where it counts.

AI / ML

  • Advanced RAG
  • Agentic Orchestration
  • Transformer Fine-tuning
  • Open-weight & Frontier LLMs
  • Multimodal
  • NLP
  • Computer Vision
  • Classical ML

Frameworks

  • PyTorch
  • Hugging Face
  • TensorFlow
  • scikit-learn
  • MLflow

Programming & Data

  • Python
  • SQL
  • PySpark
  • pandas
  • PostgreSQL

Cloud & MLOps

  • Microsoft Azure
  • Kubernetes
  • Docker
  • Argo Workflows
  • API Management
  • Jenkins
  • CI/CD

Software Engineering

  • REST APIs
  • System Architecture
  • Full-Stack Development
  • OpenWebUI
  • Flask
  • Streamlit
  • Elasticsearch

Leadership

  • Technical Leadership
  • Mentoring
  • Product & Stakeholder Management
  • Roadmaps
  • Agile Delivery

Languages

  • English (fluent)
  • German (C1)
  • Mandarin (native)

Contact

Let's build something.

Frontier AI is what gets me up in the morning. If you're building something ambitious, I'd love to hear about it.

Munich, Germany