MediPal

Your AI-powered medical assistant for overcoming all barriers to quality healthcare.

Inspiration

MediPal was born from personal travel experiences. Running out of medication in a foreign country where you don't speak the language is frustating, stressful and potentially dangerous. We wanted to build something that could bridge that gap — a tool that doesn't just translate words, but interprets medical context so you get the right medication and assistance, not just a rough translation.

What It Does

MediPal is a personal agent that acts as real-time bridge between you and a pharmacist/doctor/waiter abroad. You store your medications and health profile, and when you walk into a pharmacy or hospital, MediPal:

Interprets your medical needs to the pharmacist/doctor in their language — with voice output so you can simply hand over your phone
Surfaces active compounds, drug classes, and equivalents so they understand exactly what you need
Translates their response back to you, flagging safety concerns like dosage mismatches, drug class differences, or interactions

How We Built It

We ideated the concept together as a team, then split into focused roles: one teammate handled the presentation and project administration (scoping, slide deck, demo scripting, documentation), while the other drove the technical implementation (architecture, API integrations, data engineering).

Our stack:

Nanobot as base agentic framework (https://github.com/HKUDS/nanobot)
Telegram Bot API for a zero-install, globally accessible interface
OpenCode API for multilingual medical reasoning and interpretation
ElevenLabs for natural text-to-speech in the pharmacist's language
Python tying it all together

Challenges We Ran Into

The agent prompting beyond our scope: The agent would sometimes go into lengthy explanations that were outside the role of our interpreter. Reining the agent in to strictly interpret and surface stored profile data — without overstepping — took careful prompt engineering and iteration.
Speaking in two languages in the same prompt: Getting the model to produce clean, well-separated output in two different languages (e.g., English for the user and Thai for the pharmacist) within a single response was tricky. The languages would bleed into each other, or the model would default to one language for both sections.
Voice latency: Generating voice via the ElevenLabs API and delivering it as a Telegram voice note introduced noticeable delay in the conversation flow. For a real-time pharmacy interaction, every second counts.

What We Learned

We learned how to integrate voice capabilities with an AI agent — something neither of us had worked with before. Combining text-based LLM reasoning with speech synthesis into a coherent, real-time conversational flow was a new challenge that pushed us to think about UX beyond just text on a screen. We also gained a deeper appreciation for prompt engineering as a craft — small changes to the system prompt had outsized effects on the quality and safety of the bot's output.