Reka AI

The product

Reka’s 2023 research application let users supply their own data to fine-tune models, then explore the results through streaming conversations. It connected model customization with a practical way to examine a model’s responses.

My engagement

Beginning in March 2023, I worked as the sole client-side engineer on a greenfield AI research application. I built the client from scratch, covering assigned-token access, model fine-tuning workflows, and streaming conversations. The scope connected the user’s first access to the application with their later interaction with a customized model.

I integrated Python APIs that exposed the fine-tuning capabilities, building the user-facing path for supplying data and using the resulting models in chat. The backend services handled training and inference; my responsibility was the application through which users accessed those capabilities.

The conversational experience completed that workflow. I worked on progressive responses and conversation handling so model customization led into an ongoing exchange within the same application, rather than a disconnected set of research features.

// engineering work

• Greenfield application engineering

  I established the client application and integrated assigned-token access. I connected that access model to the research workflows, keeping client actions aligned with the operations supported by the backend contracts.

• Model fine-tuning integrations

  I integrated Python APIs that let users supply their own data for model customization. The client connected data submission and fine-tuning interactions with the resulting model’s conversational experience. This work concerned access to the training services, not implementation of the models or training infrastructure.

• Streaming conversational experiences

  I built the chat interface to display responses progressively as they arrived while maintaining the ongoing conversation. This completed the customization journey by letting users interact with their fine-tuned models in the same application.

Delivery and reflection

The application brought token-based access, model customization, and streaming conversations together. Building it strengthened my approach to greenfield engineering: define the client’s responsibilities around the user journey and validate those responsibilities against the capabilities of the underlying services.