I Thought We Were Building a Chatbot. We Were Building Something Else.

Description and use

Meena, an AI assistant for college students, was redesigned with culturally relevant UI, reranking, Gemini embeddings, web search via Tavily, bounded tool usage, and a lightweight observability layer to improve relevance, cost, and reliability.

Problem

The original AI assistant Meena was functional but lacked cultural relevance, personalized context, and robust tool‑use handling, limiting daily adoption among college students.

Outcome

Meena now provides culturally relevant, context‑aware answers with reduced noise and cost, and its architecture supports future personalization.

Workflow

• 1. Redesign UI with Indian visual culture colors and campus motifs.
• 2. Add reranking layer using Jina to filter retrieved chunks before LLM.
• 3. Switch to Gemini embeddings for richer multimodal representation.
• 4. Integrate Tavily web‑search tool for up‑to‑date information.
• 5. Implement tool‑calling boundaries: max turns, fallbacks, final synthesis.
• 6. Build lightweight observability to log model calls, retrievals, tool usage, and failures.
• 7. Plan for user‑context personalization to tailor responses per student.

Evidence

We added another layer. Reranking. Jina became part of that pipeline…

Source

We gave Meena tools… Tavily became part of the tool layer…

Source

We built our own lightweight observability layer…

Source