My Friend Was Guessing Hostel Mess Portions, So I Built Him an Offline AI Macro Agent

T Abishek — Mon, 05 Oct 2026 06:58:20 +0000

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

What I Built

If you’ve ever lived in a freshers' hostel, you know the daily struggle: you start the semester with massive fitness goals, buy a brand-new tub of whey protein, hit the gym religiously, and then walk straight into the campus mess hall to face whatever unidentifiable, liquid-gold oily curry the kitchen decided to whip up today.

The problem? Commercial fitness apps like MyFitnessPal are fundamentally built for Western grocery stores or standardised restaurant chains. Typing "hostel mess dal fry" or "random cafeteria paneer butter masala" into a generic macro tracker yields wildly inconsistent estimates-one database entry says a bowl is 100 calories, another says 450 calories, and none of them tells you how much you should actually scoop onto your tray to hit your specific targets without overshooting calories.

Watching him stare soulfully at a tray of mystery gravy every afternoon, trying to calculate mental math while the lunch queue backed up behind him, was equal parts inspiring and tragic. I decided to fix it.

Demo

Code

Abishekcps / hostelmacro

How I Built It

I built this as a local Streamlit application running on Windows. Instead of relying on a giant cloud API, I wanted the brain of this thing sitting right on his laptop, so it actually works when he needs it.

Core Intelligence (Open Weights): I used Google's open-weight Gemma 2 (gemma2:2b) running locally via Ollama. It takes the messy hostel menu, the web context, and his macro goals, and acts as the reasoning engine to build the actual plate strategy.

Web Tooling: I used the DuckDuckGo Search Python package for a keyless, zero-auth live web context. It pulls real-time nutritional profiles for specific Indian dishes before the LLM generates an answer, so it doesn't hallucinate calories.

Tracing & Performance: I integrated Sentry Agent Tracing via their Python SDK. It monitors the execution timeline, creating transaction spans for both the web tool search latency and the local LLM generation time, so I can see exactly where the bottleneck is if the app feels slow.

Why Does Open Innovation Matter?

Why didn't I wrap a closed cloud API in a website? Two words: Cost and Connectivity.

When you are checking your macros three times a day, 30 days a month, API token bills add up fast for a broke college student. By using an open-weight model like Gemma 2, this app costs absolutely $0 to run forever.

Secondly, hostel Wi-Fi can be notoriously spotty, especially during peak hours. Building this with local inference means the core intelligence works even when the internet drops completely. Furthermore, his personal health goals and daily dietary logs remain 100% private on his own hardware. Open source means anyone reading this can clone the repo, run ollama pull gemma2:2b, and have a fully functional AI assistant running on their machine tonight without setting up credit cards or secret API keys.

Prize Categories:

Best Use of Gemma: Running Google's open-weight gemma2:2b locally via Ollama as the core reasoning engine for macro calculations and meal planning.

Best Use of Sentry Agent Tracing: Full telemetry integration tracking tool execution spans, error logging, and latency profiling across the local LLM inference and web context tools.

DEV Community: T Abishek