AI enthusiast Muhammad Zulali managed to configure his 32-year-old 1994 Jeep Cherokee SUV to pass strict California emissions tests by using the Claude and Codex AI models and an electronics kit costing less than $15.

Problem and costs of blind adjustments

The old SUV failed environmental tests due to elevated levels of nitrogen oxides at certain speeds. Initially, the car failed the test at a speed of 25 mph (about 40 km/h), and after manual attempts by Zulali to adjust the fuel-air mixture, the situation only worsened — the readings exceeded the norm at both 25 and 50 mph (80 km/h), writes Tom's Hardware.

Since he had to act blindly, and each re-check cost almost $100, the car owner decided to connect artificial intelligence and a $3 microcontroller to the diagnostics (the total cost of all components was no more than $15).

How AI and Arduino were connected to a 1994 car

The main obstacle was the incompatibility of a modern laptop and the primitive electronic control unit of a 1994 car. With Claude's hints, it turned out that the engine uses a Chrysler SCI bus with a data rate of about 7800 baud and an electrically inverted signal.

To solve this problem, an enthusiast assembled a circuit on a breadboard with an Arduino Nano microcontroller, connecting it with three wires to the Jeep's 6-pin diagnostic connector. After a successful connection, the system initially read only the coolant temperature and battery voltage, but Zulali later configured the collection of extended telemetry and error codes during test drives at speeds of 25 and 50 mph, and also gained the ability to control parameters such as fuel pressure.

Bypassing memory limits and open source

When the cause of the malfunction was found, it turned out that flashing the main ECU was impossible — the car's computer turned out to be several months older than the models that supported firmware updates. Then Zulali brought in Codex, which helped implement and run custom code through an adjacent EEPROM memory chip.

"Claude performed most of the diagnostic work, and Codex did almost all of the memory work. I just connected the wires and drove the car," Zulali noted, adding that he was able to do without the services of mechanics and expensive Chrysler dealership equipment.

The enthusiast released their developments and diagnostic code publicly on GitHub under the name xj-scanner under the GPLv3 license.