Where LLMs Have Been Incredibly Useful
Budget-Friendly Engineering Hacks
One of the biggest challenges in Lemons racing is the $500 car budget cap. Very early analysis offered some guidance on cars that might have valuable parts because their are communities that still drive and maintain them. And there might also be some harder to find parts like pop-up light motors or interior switches and gauges. Then there is parts substitution – i.e. are their cheap versions of parts like coilovers or injectors. This helped guide us to 944/924.
As we got into specifics, AI proved surprisingly valuable at finding cheap substitutes for expensive parts:
- Swapping in a BMW transmission cooler as an oil cooler, found for just $30 on eBay, instead of buying specialized equipment. Update – In the end, with specialized fittings, it turned out that an Alibaba unit was cheaper.
- Identifying that a Vanagon fuel vapor canister could replace our failing Porsche part—for $19 instead of $360.
- Recommending Superbeetle coilovers and Golf coilovers to upgrade suspension, delivering race-worthy handling at a fraction of the cost. About 10x cheaper than other options we found.
These suggestions, scraped from years of forum discussions, turned out to be not just viable but highly effective.
Troubleshooting & Guidance
LLMs excelled at helping us cut through the noise of old forum posts and manuals. Some highlights:
- Translating our string-method suspension alignment measurements into specific tie rod adjustments that worked remarkably well.
- Quickly generating multiple options for upgrades, which we then cross-checked against human expertise (“trust but verify” became our mantra).
- Accelerating problem-solving by synthesizing scattered knowledge into concise recommendations.
On-Track Performance
The proof was in the lap times. Thanks to AI-assisted preparation:
- The car handled sustained high temperatures with solid thermal management.
- Suspension and alignment tweaks made the car feel dialed in from day one, rather than after endless trial-and-error.
We’ll do another update in 10 days after a lot of safety work, so hopefully we didn’t break anything.
Where LLMs Still Struggle
Innovation & Novel Design
AI is great at recombining existing knowledge, but not at inventing. When we asked for aerodynamic ideas, we got the same wing and splitter setups you’d find on any forum thread. Even small adjustments had mixed results.
One emerging idea from one of our founders is the idea that LLMs need a good digital twin to test against. i.e. you don’t need LLMs to model everything, but you do need models. So for example, if you have a model for piece of aero like a front splitter, you can do some optimization by having LLMs explore and compare resulting performance of the model with different inputs.
Price Accuracy & Trade-offs
AI often nailed the what, but not always the how much. For example, with a faulty MAF sensor, LLMs suggested both DIY builds and Alibaba drop-in replacements. Sorting through cost, reliability, and packaging trade-offs still required a lot of human judgment. In the end the parts list for a clever hack to adapt a modern MAF to output voltage in the range suitable for the 924 engine management system…proved more expensive than sourcing a single replacement part which also saved time.
Visualization & Spatial Reasoning
Ask an LLM for a diagram, and the results can be… let’s just say confusing. Numerical advice for suspension adjustments was spot on, but when it came to 3D visualizations, we were left scratching our heads.
Human Validation Still Essential
At the end of the day, every promising suggestion still had to be double-checked. AI accelerated the search, but the community of human mechanics, racers, and tinkerers remained essential for confirmation.
The Takeaway
AI has proven to be a powerful budget-conscious co-pilot. It’s helped us move faster, cheaper, and sometimes smarter. But it’s not magic. The best results came when we treated LLMs as idea generators and research accelerators, and then layered in human expertise for validation, creativity, and final decision-making. This is inline with what we’re seeing from successful startup adoption, especially in hardware engineering.