Batu Thoms, product development, environmental and AI engineer

Engineering for AI and sustainability.

I design and build products, measure how they impact the environment, and build tools that automate and optimise workflows.

Portrait of Batu Thoms
Based in
Odense, Denmark
Studying
MSc Environmental Engineering, SDU
Graduating
June 2027
Speaks
German, English

About

Hi, I’m Batu, an engineer in Odense. I started in product development and moved to environmental engineering to understand the impact our products have on the environment.

Today I work as a student research assistant at SDU, implementing AI in sustainability, and I am building AI into my own skill set. I graduate in June 2027.

Profile

Skills and tools

I know how to design and build products.

Three years of product development at SDU, and a thesis with AVK on sensors in water valves.

I am learning to measure their impact, responsibly.

Life cycle assessment, energy systems and green fuels, in my master’s at SDU.

Now I am working out where AI can take over the steps.

Retrieval pipelines and AI agents that do the reading and checking, and show their sources.

Build, measure, automate.

How a thing is made, how it impacts the environment, and data you can trust about both.

The full list

Environment

Life cycle assessment

  • openLCA
  • ecoinvent
  • PEF

Used inRecycling wind turbine bladesProduct environmental footprint for CustomiseMe

Energy and cost modelling

  • EnergyPLAN
  • Excel
  • Power-to-X

Used inThe green methane cost model.

Carbon accounting and ESG

  • GHG Protocol
  • ESG reporting

Used inResearch assistant work at SDU.

Maps and material flows

  • QGIS
  • Material flow analysis

Used inThe building stock of Odense.

AI and data

Retrieval and RAG

  • Python
  • Embeddings
  • Vector search
  • Chunking
  • PDF parsing

Used inThe RAG pipeline for sustainability reports.

Evaluating AI systems

  • Test sets
  • Scoring scripts
  • Error analysis
  • Regression checks

Used inEvery change to the RAG pipeline was kept or dropped on its score.

AI agents and automation

  • Agent harnesses (Claude Code)
  • Guardrails
  • Hooks
  • Subagents
  • Skills
  • Prompt and context design

Used inA notebook kept by AI agents.

Local and private AI

  • Ollama
  • Local LLMs
  • Data handling rules

Used inMy own daily setup, built around confidential research data.

Data analysis

  • Python
  • pandas
  • matplotlib

Used inCoursework at SDU and my research assistant work.

Product

Product development

  • CAD modelling
  • 3D printing
  • Prototyping
  • User interviews

Used inSensors inside water valves, with AVK.

Project management

  • Planning and scheduling
  • Risk and stakeholders
  • Microsoft Office

Used inGroup projects in every semester since 2022.

Still learning: SQL, Git and Danish.

How I work

  1. Problem first

    Find the real problem, then pick the tool.

  2. Prototype before product

    A small working version before anything big.

  3. Check, then trust

    Results get tested before I rely on them.

  4. AI as a teammate

    It earns more responsibility as it proves itself.

  5. Team over solo

    Goals and deadlines fit the team, not just me.

Selected work

Illustration: an open report with one passage highlighted and lifted out onto a small glass panel
Research at SDU Green Technology2026

A RAG pipeline for sustainability reports

A retrieval pipeline that reads suppliers’ sustainability reports and returns each figure with the passage it came from. I designed the framework and the test that scores it.

  • Python
  • RAG
  • Embeddings
  • PDF parsing
  • Evaluation
  1. Parse: Page by page, text and tables.
  2. Split and embed: Short passages, each a vector.
  3. Retrieve: The closest passages per question.
  4. Read the value: Written rules pick the figure.
  5. Keep the evidence: Stored with its source passage.
  6. Score and repeat: Each run is compared with answers checked by hand.
Close-up of the AVK logo cast into a blue valve body
Bachelor thesis with AVK2025

Sensors inside water valves

Wireless pressure sensors fitted into industrial gate valves, tested on production hardware.

  • CAD
  • IoT
  • Prototyping
  • Interviews
  1. Interview the product teams and write down the requirements.
  2. Design integration concepts in CAD.
  3. Test them hands-on with production valves and sensors.
  4. Work out the business case for predicting maintenance.
CustomiseMe logo
MSc in-company project2026

Product environmental footprint for CustomiseMe

An in-house life cycle assessment of the company’s products.

  • openLCA
  • Life cycle assessment
  1. Agree the goal, the functional unit and the system boundary with the company.
  2. Collect the inventory data from the company.
  3. Model the products in openLCA and calculate the impacts.
  4. Show where the footprint comes from and what would lower it.
Wind turbine blades stored on an industrial site
MSc project2025

Recycling wind turbine blades

A life cycle assessment comparing ways to recycle old blades.

  • openLCA
  • Life cycle assessment
  1. Set the goal, the functional unit and the system boundary.
  2. Build the inventory for each recycling route in openLCA.
  3. Calculate the environmental impacts.
  4. Compare the routes and state what the result depends on.
A methanation plant in Allendorf (Eder), Germany: reactor, pipework and container units
MSc project2025

Can green methane compete?

A cost model of a renewable methane plant against a fossil fuel.

  • Power-to-X
  • Cost model
  1. Define the plant and the fossil fuel it is compared with.
  2. Build the cost model: investment, running costs, output.
  3. Vary the electricity price and the running hours.
  4. State under which conditions the green fuel competes.
Diagram: an LLM wiki in the middle, connected to Obsidian, Git, Markdown, a terminal, a browser clipper and an AI model
Personal2026

A notebook kept by AI agents

A knowledge base kept by AI agents that follow written rules and show what they checked.

  • Agent harness
  • Agents
  • Obsidian
  1. Write the rules down in plain language.
  2. Let scripts enforce the rules that a script can check.
  3. Hand each task to an agent sized for it.
  4. End every large task with a list of what was checked and what was not.
Illustration: a friendly robot in a graduation cap behind an open book, saying Hej, in front of the Danish flag
Personal2026

Danish tutor

We have Duolingo at home.

  • Claude Code plugin
  • Spaced repetition
  • SQLite

More code on GitHub

Background

  1. 2022 to 2025

    BSc Product Development and Innovation Engineering

    SDU.

  2. 2024 to 2025

    Exchange semester, FH Kufstein Tirol

    Product Engineering and Management, Austria.

  3. 2025

    Bachelor thesis with AVK

    Wireless pressure sensors in industrial valves.

  4. 2025 to 2027

    MSc Environmental Engineering

    SDU Odense. Graduating June 2027.

  5. 2026 to now

    Student research assistant

    SDU Department of Green Technology.

Want to get in contact with me?