Overview
NI leadership understood the potential of LLMs, but knew that applying these generic models to technical use cases would require innovation and experimentation. I was part of a small team focused on identifying internal use cases and designing, prototyping, and testing AI and LLM powered solutions.
One area we focused on was ADAS (Advanced Driver Assistance Systems) camera validation. The existing process was very manual, taking internal ADAS engineers weeks to solve, creating a bottleneck to serving more customers.
My contribution
Co-design facilitation
UX design
UI design
The team
3 x full-stack developers
1 x data scientist
Tools
LucidChart Paper & pen Figma
Process
Weekly sessions with users
Our innovation team met weekly with NI's ADAS directors and engineers. At first we aimed to understand their work at a high level, then narrowed our focus to the task of sensor and bus emulation as it became clear that was an area where we could likely leverage AI and LLM technologies to drive efficiencies.
UX Mission to drive conceptual discussions
An early co-design session with the innovation team and ADAS engineers produced detailed ideas around facilitating the use of sensor documentation (usually in pdf form and searched manually) and small improvements to existing spreadsheets. I saw an opportunity for the team to dream bigger and created a UX Mission deck to refocus our discussions around a higher-level concept, knowing that we could iterate on the details if we could nail the concept.
Early experiments boost confidence
Our ADAS engineers found the concept intriguing, but, because it was based on a novel and untested approach of using AI to find patterns in register and I2C data, the team was hesitant to continue in this direction until we had evidence it would even be possible. A week after reviewing the concept, our team's data scientist built and demo'd a ML powered solution that was more than 98% accurate in identifying relationships between register changes and I2C transactions - something that took ADAS engineers days to solve manually.
Outcome
The early ML-backed prototype exceeded expectations, identifying at least 98% of the register-I2C relationships in the test data. With this increased confidence, the team continued to build in this direction, demoing weekly to a growing audience. I deepened my collaboration with the developer building the front-end, ensuring we used modified base components (Material Design) and a programmatic color scheme (modified Flexoki scheme).
This work also helped NI leadership see the value in co-design and early and frequent prototyping, especially when building with emerging technologies. It was included in case-studies used to define best practices when building with LLMs for NI's R&D Directors.
The team was still iterating and building in Nov 2023 when I was part of sweeping layoffs following Emerson's acquisition of NI.