Published on July 1, 2025

Last week, I took part in my very first hackathon - the Bosch Agentic AI Hackathon in Berlin - and what an incredible two days it turned out to be. Not only did I have the chance to immerse myself in the rapidly evolving space of Agentic AI, but my team (composed of myself, Daniel Glatter, Karolina Kur & Andreas Spannagel) also ended up taking first place! 🥇 We built an autonomous agent that tackles a very real industry pain point: optimising PCBA (Printed Circuit Board Assembly) Bill of Materials (BOM) files. Think of it as an AI co-pilot for hardware procurement.

Our project aimed to solve a common problem in electronics manufacturing - how to analyse and optimise a BOM file quickly and intelligently. BOM files often contain dozens or hundreds of components which need to be procured, and finding the most cost-effective ones on the market can be time-consuming and error-prone. We built a multi-agent system that does the heavy lifting automatically. Here's how it works:

1. Input Processing: When a user uploads a BOM file, the agent initiates a conversation to gather critical project assumptions (industry, order quantity etc.).
2. Part Analysis: It then parses the BOM using a Python tool to: identify key columns via an LLM, structure each row with LLM help and query Octopart for data on original parts and alternatives.
3. Alternative Evaluation: For each part with alternatives, the agent checks if substitutions are suitable using LLM-based reasoning.
4. Report Generation: The agent generates a markdown report summarising recommendations, justifications, and potential cost savings.

Code for full agent workflow available open access on Github

Agentic AI goes beyond traditional automation. It's about giving systems autonomy to reason, plan, and act within a bounded goal. The system has both a brain and tools to use for action. This approach made perfect sense for our use case, where we needed to:

• Chain multiple reasoning steps across structured data
• Adapt to incomplete or ambiguous inputs
• Respond to user prompts dynamically with new results

Key Takeaways

Reflecting on the Hackathon, I walked away with more than just a prize:

• Make that first move! Agentic AI was relatively new to me, and I was nervous at first, knowing nobody at the event. But jumping in turned out to be the best way to learn and I was able to contribue more than I expected!
• Code Fast, Debug Slow Rapid, vibe coding and prototyping is exhilarating, but when the bugs hit (and they will), clear thinking and good logging save the day.
• AI as a Creative Partner Agentic systems aren't just efficient, they're surprisingly imaginative. They can spot patterns and connections across tasks that humans might overlook.
• Teamwork makes the difference By first understanding each other, our skillsets and expectations, we could lay a strong base for the hacking.

What's next?

I’m more excited than ever about the real-world applications of Agentic AI, not just in procurement, but in domains like logistics, legal research, weather forecasting and even design. I’m already thinking about how to bring agentic workflows into future projects and (hopefully!) more hackathons.
Huge thanks to Bosch Digital for organising such a thoughtful and well-run event, and to all the partners — Microsoft, AWS, NVIDIA, Capgemini, n8n, KI Park & IBM — for the support.

• © Untitled
• Design: HTML5 UP