The Shape of Code

This weekend, Team Awesome (Sam, Frank and yours truly) took part in the [London] ClearRoute x Le Mans 24h Hackathon 2025 (ClearRoute is an engineering consultancy and Le Mans is an endurance-focused sports car race).

London hackathons have been thin on the ground during the last four years. I suspect that the chilling of the economic climate, with the end of the zero interest-rate policy, caused companies to cut back funding for projects whose benefits were rather indirect. Things do seem to be picking up. This is my second hackathon this year, there are two hacks next weekend and one the following weekend.

Based on the title, the theme of the hackathon was obviously the Le Mans 24 hour car race, and we were asked to use ClearRoute’s LLM-based tools to find ways to improve race team performance.

I was expecting the organisers to provide us with interesting race data. After asking about data and hearing the dreaded suggestion, “find some on the internet”, I was almost ready to leave. However, the weekend was rescued by a sudden inspired idea.

My limited knowledge of motorsport racing comes from watching Formula 1 on TV (until the ever-increasing number of regulations created a boring precession), and I remembered seeing teams penalized because they broke an important rule. The rule infringement may have been spotted by a race marshal, or a member of another team, who then reported it to the marshals.

Le Mans attracts 60+ racecars each year, in three categories (each with their own rules document). The numbers for 2025 were 21 Hypercars, 17 LMP2 prototypes, and 24 LMGT3 cars (the 2025 race ran this weekend).

Manually checking the behavior of 60+ cars against a large collection of ever-changing rules is not practical. Having an LLM-based Agent check racing events for rule violations would not only be very cost-effective, but it would also reduce the randomness of somebody happening to be in the right place and time to see an infringement.

This idea now seems obvious, given my past use of LLMs to check software conformance and test generation.

Calling an idea inspired is all well and good, if it works. This being a hackathon, suck-it-and-see is the default response to will it work questions.

One of the LLMs made available was Gemini Flash, which has a 1 million token input context window. The 161 page pdf of the Le Mans base technical rules document probably contains a lot less than 1 million tokens. The fact that the documents were written in French (left column of page) and English (right column) was initially more of a concern.

Each team was given a $100 budget to spend on LLMs, and after spending a few percent of our budget we had something that looked like it worked, i.e., it detected all 14 instances of race-time checkable rule violations listed by Grok.

My fellow team-mates knew as much about motor racing as I did, and we leaned heavily on what our favourite LLMs told us. I was surprised at how smoothly and quickly the app was up and running; perhaps because so much of the code was LLM generated. Given how flawed human written hackathon code can be, I cannot criticize LLM generated hackathon code.

Based on our LLM usage costs during application creation and testing, checking the events associated with one car over 24 hours is estimated to be around $36.00, and with a field of 60 cars the total estimated cost is $2,160.

Five teams presented on Sunday afternoon, and Team Awesome won! The source code is available on GitHub.

Motorcar racing is a Red Queen activity. If they are not already doing so, I expect that teams will soon be using LLMs to check what other teams are doing.

Thanks to our ClearRoute hosts who kept us fed and watered, and were very responsive to requests for help.