- Client: Bebis (Pasifik Elektrik)
- Role: Senior technical partner
- Engagement: September 2023 – present
- Stack: Speech recognition, LLM disambiguation, SQL
Bebis (Pasifik Elektrik) operate one of Turkey’s larger nutrition and diet-planning platforms. Among other projects I have worked on for them, they asked me to help as a senior technical partner for a voice-driven meal-logging prototype. Users can describe a meal in natural Turkish, such as “a glass of milk, two slices of bread, half an apple”. After that, the system then matches it to the food and portion database, with the correct nutritional weights.
The work has included:
- End-to-end pipeline design for the voice-logging prototype, whichcovers speech recognition, language-model disambiguation, and portion-weight matching against the Bebis food database.
- A database-quality review based on testing by native Turkish speakers. About half of the failed matches came from data problems, not algorithm problems. As a result, I wrote a detailed issues report, led meetings with the dietitian and developers, and proposed schema changes. These alias tables and language-coded data sources will make future versions of the software faster, cheaper to run and more accurate.
- An internationalisation strategy memo for Bebis’ planned multi-market expansion.
- Planning for the production rebuild that will follow the data cleanup.
The engagement combines product design, technical judgement, hands-on prototyping, and SQL. I get involved wherever that combination is useful, but most of the final build will be done by other developers.