Luciq · AI features
AI features inside the Luciq dashboard
A case study from Bola Sydhom’s work at Luciq. Customers wanted AI help inside the dashboard. But LLM answers are slow and sometimes wrong, and engineers rely on this tool every day. The AI features shipped, and when the model fails, users still get a working screen instead of an error. We kept improving them in short cycles, based on what customers told us.
The problem
Customers wanted AI help inside the dashboard. But LLM answers are slow and sometimes wrong, and engineers rely on this tool every day.
What I did
- Connected LLM APIs through a Node.js service
- Streamed responses into the Vue.js UI, so answers show up as they’re written
- Added checks and fallbacks for when the model gets it wrong
- Tested answers for accuracy before release, then tuned prompts from customer feedback
The outcome
The AI features shipped, and when the model fails, users still get a working screen instead of an error. We kept improving them in short cycles, based on what customers told us.
Skills used
- LLM APIs
- Node.js
- Streaming
- Vue.js
- Prompt tuning
How I think about it.
- AI helps, people decide
- Masroof’s receipt scanner fills in items for you to check, and past purchases override the model. At Luciq I added checks and fallbacks for unreliable LLM output.
- Watch real use, then improve
- When iPhone barcode scanning didn’t work well, I swapped the scanner and added typing the code in. At Luciq, we tuned prompts from real customer feedback.