<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Arman Sinha</title>
    <description>The latest articles on DEV Community by Arman Sinha (@arman_sinha).</description>
    <link>https://dev.to/arman_sinha</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4153680%2F15652e80-14fb-4489-9fca-ecbf9ebd5d1d.jpg</url>
      <title>DEV Community: Arman Sinha</title>
      <link>https://dev.to/arman_sinha</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/arman_sinha"/>
    <language>en</language>
    <item>
      <title>PHANTOM ROOM: Giving a Room's Sensor Data a Ghost Story with Local AI</title>
      <dc:creator>Arman Sinha</dc:creator>
      <pubDate>Fri, 30 Oct 2026 20:32:45 +0000</pubDate>
      <link>https://dev.to/arman_sinha/phantom-room-giving-a-rooms-sensor-data-a-ghost-story-with-local-ai-6ja</link>
      <guid>https://dev.to/arman_sinha/phantom-room-giving-a-rooms-sensor-data-a-ghost-story-with-local-ai-6ja</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-2024"&gt;Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machines&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;What happens in your room when nobody is watching?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;At 2:14 AM, the temperature suddenly drops.&lt;/p&gt;

&lt;p&gt;At 2:16 AM, the distance sensor records an unexpected change.&lt;/p&gt;

&lt;p&gt;At 2:18 AM, another sensor picks up an unusual reading.&lt;/p&gt;

&lt;p&gt;What if your room could tell you its own story?&lt;/p&gt;

&lt;p&gt;That's the idea behind &lt;strong&gt;PHANTOM ROOM&lt;/strong&gt;, an AI-powered environmental monitoring system that transforms sensor readings into mysterious, atmospheric narratives using locally running open-weight AI.&lt;/p&gt;

&lt;p&gt;Rather than simply displaying numbers and graphs, Phantom Room gives environmental data a voice.&lt;/p&gt;

&lt;p&gt;It combines an ESP32 with three sensor modules:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;DHT11:&lt;/strong&gt; Temperature and humidity monitoring.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;HC-SR04:&lt;/strong&gt; Ultrasonic distance monitoring.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;MQ-2:&lt;/strong&gt; Raw gas and smoke sensor readings.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The application continuously collects readings, identifies unusual patterns and records events in a chronological timeline.&lt;/p&gt;

&lt;p&gt;But here's where it gets interesting.&lt;/p&gt;

&lt;p&gt;A local AI model takes those observations and turns them into eerie investigative stories, supported by actual sensor evidence.&lt;/p&gt;

&lt;h3&gt;
  
  
  Some of its key features:
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Five simulation scenarios:&lt;/strong&gt; Normal Night, Temperature Anomaly, Unknown Movement, Gas Anomaly and Phantom Sequence.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Environmental anomaly detection:&lt;/strong&gt; Rolling baselines, configurable thresholds and event cooldowns.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Multi-sensor anomaly detection:&lt;/strong&gt; Identifies multiple anomalies occurring in the same sample.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Local AI storytelling:&lt;/strong&gt; Generates atmospheric narratives using Ollama and an open-weight language model.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Offline fallback:&lt;/strong&gt; A rule-based narrative generator keeps the application functional when Ollama is unavailable.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Investigation dashboard:&lt;/strong&gt; Live sensor cards, telemetry graphs, event timelines and a Phantom Intensity Score.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Session recording and replay:&lt;/strong&gt; Preserves historical readings and events, with CSV and JSON exports.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The application also supports a complete simulation mode, so the experience can be demonstrated without connecting physical hardware.&lt;/p&gt;

&lt;p&gt;And no, it doesn't actually detect ghosts. It detects environmental anomalies and gives them a fictional narrative interpretation.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;Phantom Room currently runs locally rather than on a publicly deployed server.&lt;/p&gt;

&lt;p&gt;You can launch it using the instructions in the repository:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/ArmanSinha7/phantom-room#quick-start-guide" rel="noopener noreferrer"&gt;&lt;strong&gt;Run Phantom Room — Quick Start Guide&lt;/strong&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;For the most interesting demonstration:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Start the application.&lt;/li&gt;
&lt;li&gt;Select the Phantom Sequence scenario.&lt;/li&gt;
&lt;li&gt;Watch the sensor readings and anomaly timeline change.&lt;/li&gt;
&lt;li&gt;Generate a story using the local AI engine.&lt;/li&gt;
&lt;li&gt;Explore the recorded session and its sensor evidence.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The repository also includes a hackathon demo walkthrough.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;

&lt;p&gt;The complete source code, ESP32 firmware, documentation and implementation specification are available here.&lt;/p&gt;


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/ArmanSinha7" rel="noopener noreferrer"&gt;
        ArmanSinha7
      &lt;/a&gt; / &lt;a href="https://github.com/ArmanSinha7/phantom-room" rel="noopener noreferrer"&gt;
        phantom-room
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;🔮 PHANTOM ROOM&lt;/h1&gt;
&lt;/div&gt;
&lt;div class="markdown-heading"&gt;
&lt;h3 class="heading-element"&gt;Environmental Intelligence System &amp;amp; Haunted Room Simulator&lt;/h3&gt;
&lt;/div&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;"Sometimes, the room remembers things we don't. Phantom Room gives those silent signals a voice."&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Built for the &lt;strong&gt;Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine&lt;/strong&gt;.&lt;/p&gt;
&lt;p&gt;&lt;a href="https://github.com/ArmanSinha7/phantom-room/./LICENSE" rel="noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/9e04966f6a646b1cd3fc05c6ea88a80211c6f20b186b584c2c515ef9c411ad2c/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4c6963656e73652d4d49542d616d6265722e737667" alt="License: MIT"&gt;&lt;/a&gt;
&lt;a href="https://ollama.com" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/a82e6f4f34ac330f492e0e34180fafe07293aa57fd4e4c52367c5326e64d674a/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4c6f63616c5f41492d4f6c6c616d615f4f70656e5f576569676874732d3130623938312e737667" alt="Local AI: Ollama"&gt;&lt;/a&gt;
&lt;a href="https://fastapi.tiangolo.com" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/5722dd0b157f7b4d3434ce322fb68a2029c6eab6f587773f03f77aed4eb6c12c/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4261636b656e642d466173744150495f2532425f53514c6974652d3030393638382e737667" alt="Backend: FastAPI + SQLAlchemy"&gt;&lt;/a&gt;
&lt;a href="https://react.dev" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/3d5fec15ad53fdcc5e3b9fe043e6ace75cd72aa15468f774d77f264fd49b11c8/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f46726f6e74656e642d52656163745f2532425f547970655363726970745f2532425f5461696c77696e642d3338626466382e737667" alt="Frontend: React + TypeScript"&gt;&lt;/a&gt;
&lt;a href="https://www.espressif.com" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/eabdad813e62dee3b6f83d52beac384171f14b7ffc6acd5d7456cc5ff2a97bec/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f48617264776172652d45535033325f4465764b69742d4537333532432e737667" alt="Hardware: ESP32 DevKit"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;1. Concept &amp;amp; Problem Statement&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;What happens in your room when nobody is watching?&lt;/p&gt;
&lt;p&gt;Traditional IoT environmental monitors display sterile, clinical numbers—&lt;code&gt;22.4 °C&lt;/code&gt;, &lt;code&gt;48% RH&lt;/code&gt;, &lt;code&gt;115 cm&lt;/code&gt;, &lt;code&gt;320 ADC&lt;/code&gt;. Human observers are left to squint at graphs to notice if a sudden cold snap preceded an unusual distance deflection and a chemical spike.&lt;/p&gt;
&lt;p&gt;Conversely, ghost hunting mobile apps fabricate random spooky sounds without grounding their experience in actual physical reality.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Phantom Room bridges real physical sensing and local open-weight AI:&lt;/strong&gt;&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;It streams real environmental readings from an &lt;strong&gt;ESP32&lt;/strong&gt; (or its built-in &lt;strong&gt;5-scenario simulation engine&lt;/strong&gt;).&lt;/li&gt;
&lt;li&gt;A statistical &lt;strong&gt;Event Detection Engine&lt;/strong&gt; monitors rolling baselines, detects single-sensor anomalies…&lt;/li&gt;
&lt;/ul&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/ArmanSinha7/phantom-room" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;&lt;strong&gt;GitHub:&lt;/strong&gt; &lt;a href="https://github.com/ArmanSinha7/phantom-room" rel="noopener noreferrer"&gt;https://github.com/ArmanSinha7/phantom-room&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The project is released under the MIT License.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;I wanted the project to combine three things: IoT, open-source AI and an unusual user experience.&lt;/p&gt;

&lt;h3&gt;
  
  
  1. The hardware layer
&lt;/h3&gt;

&lt;p&gt;An ESP32 collects readings from the DHT11, HC-SR04 and MQ-2 sensors.&lt;/p&gt;

&lt;p&gt;It transmits the readings to a FastAPI backend over the local network using HTTP.&lt;/p&gt;

&lt;p&gt;I also built a simulation engine with five predefined scenarios, allowing the entire application to work without physical hardware.&lt;/p&gt;

&lt;p&gt;Both simulated and hardware readings enter the backend processing pipeline.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. The event detection engine
&lt;/h3&gt;

&lt;p&gt;I implemented a Python-based anomaly detection engine.&lt;/p&gt;

&lt;p&gt;It maintains a rolling window of up to 50 readings and calculates sensor baselines.&lt;/p&gt;

&lt;p&gt;It compares incoming readings against configurable thresholds, identifies anomalies and uses cooldown periods to prevent repetitive alerts.&lt;/p&gt;

&lt;p&gt;When multiple sensors cross their thresholds in the same sample, the system can generate a combined anomaly event.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. The AI layer
&lt;/h3&gt;

&lt;p&gt;This is the heart of Phantom Room.&lt;/p&gt;

&lt;p&gt;I integrated &lt;strong&gt;Ollama with Qwen 2.5 3B as the default model&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;Instead of sending data to an external AI API, the application sends structured sensor observations to a locally running language model.&lt;/p&gt;

&lt;p&gt;The model generates a structured narrative containing:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;A mysterious title.&lt;/li&gt;
&lt;li&gt;An atmospheric story.&lt;/li&gt;
&lt;li&gt;A chronological event sequence.&lt;/li&gt;
&lt;li&gt;Supporting sensor evidence.&lt;/li&gt;
&lt;li&gt;An eerie closing line.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The generated narrative must respect the actual observations rather than invent sensor readings.&lt;/p&gt;

&lt;p&gt;When the local model is unavailable, a Python-based rule-driven fallback generator takes over.&lt;/p&gt;

&lt;h3&gt;
  
  
  4. The application
&lt;/h3&gt;

&lt;p&gt;The frontend uses React, TypeScript, Vite, Tailwind CSS, Recharts and Lucide React.&lt;/p&gt;

&lt;p&gt;The backend uses Python, FastAPI, SQLAlchemy and SQLite.&lt;/p&gt;

&lt;p&gt;The dashboard refreshes sensor information periodically and presents it through a cinematic environmental monitoring interface.&lt;/p&gt;

&lt;p&gt;I also implemented session recording, historical replay and downloadable reports.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;This project made me think about something beyond the Halloween theme.&lt;/p&gt;

&lt;p&gt;Environmental data can be surprisingly personal.&lt;/p&gt;

&lt;p&gt;A room's temperature patterns, sensor activity and monitoring history are data that I believe users should retain control over.&lt;/p&gt;

&lt;p&gt;Using open-weight AI and local inference makes that possible.&lt;/p&gt;

&lt;p&gt;Here is what the open-based approach provides:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Privacy by design&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Ollama runs the language model locally. Sensor readings are stored in a local SQLite database instead of being sent to a third-party AI API.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Offline capability&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Once dependencies and the language model have been installed, the application can perform its monitoring, simulation and AI inference without requiring an internet connection.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. No paid AI API dependency&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;There are no per-request cloud AI API charges or mandatory external AI subscriptions.&lt;/p&gt;

&lt;p&gt;The model runs using the user's own hardware resources.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Freedom to experiment&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The default model is configurable. The application can be adapted to use other compatible Ollama models without rebuilding the entire application around a proprietary API.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. Ownership and extensibility&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The data, backend and AI integration remain under the user's control. Developers can modify the event detection logic, change the model, add sensors or create entirely new scenarios.&lt;/p&gt;

&lt;p&gt;A closed API could generate a similar story, but local inference makes privacy, offline operation and model control fundamental parts of the experience rather than optional features.&lt;/p&gt;

&lt;p&gt;For Phantom Room, that distinction matters.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Spirit of It
&lt;/h2&gt;

&lt;p&gt;The most interesting part of this project isn't the generated horror story itself.&lt;/p&gt;

&lt;p&gt;It's the idea that ordinary environmental readings, when observed together, can become a narrative.&lt;/p&gt;

&lt;p&gt;A temperature fluctuation is just a measurement.&lt;/p&gt;

&lt;p&gt;A sudden distance change is just another measurement.&lt;/p&gt;

&lt;p&gt;But when these observations occur together, they create a sequence worth investigating.&lt;/p&gt;

&lt;p&gt;I wanted to build something that makes people curious about the invisible activity around them, while keeping the underlying evidence accessible and transparent.&lt;/p&gt;

&lt;p&gt;No artificial supernatural detections. No pretending that fiction is science.&lt;/p&gt;

&lt;p&gt;Just real sensor observations, open AI and a little imagination.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Sometimes, the room remembers things we don't. Phantom Room gives those silent signals a voice.&lt;/strong&gt;&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>ai</category>
      <category>iot</category>
    </item>
    <item>
      <title>NutriGhar: A Personal Nutrition Companion Built for My Sister and Best Friend</title>
      <dc:creator>Arman Sinha</dc:creator>
      <pubDate>Sun, 04 Oct 2026 22:43:55 +0000</pubDate>
      <link>https://dev.to/arman_sinha/nutrighar-a-personal-nutrition-companion-built-for-my-sister-and-best-friend-2ed5</link>
      <guid>https://dev.to/arman_sinha/nutrighar-a-personal-nutrition-companion-built-for-my-sister-and-best-friend-2ed5</guid>
      <description>&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;I built &lt;strong&gt;NutriGhar&lt;/strong&gt; for my sister and best friend, &lt;strong&gt;Ayushi Sinha&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;Ayushi wanted to lose some weight, get fitter, and build a little more shape, and she asked me if I could make something that would help her keep track of her daily calories and protein without making the process complicated.&lt;/p&gt;

&lt;p&gt;That conversation became NutriGhar.&lt;/p&gt;

&lt;p&gt;NutriGhar is an offline-first nutrition and weight-management app designed around Indian food. It lets a user track meals, calories, protein, weight, goals, activity, and progress while working with foods and portions that actually make sense in an Indian household.&lt;/p&gt;

&lt;p&gt;For this challenge, I extended NutriGhar with an open-source AI layer that can understand natural meal descriptions and convert them into structured food and portion information before passing them to the application's nutrition engine.&lt;/p&gt;

&lt;p&gt;The idea is simple:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;You describe what you ate naturally. NutriGhar does the tracking.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;


&lt;div class="crayons-card c-embed text-styles text-styles--secondary"&gt;
    &lt;div class="c-embed__content"&gt;
      &lt;div class="c-embed__body"&gt;
        &lt;h2 class="fs-xl lh-tight"&gt;
          &lt;a href="https://drive.google.com/file/d/1UA8j76lkXg4n02VBHjSK_c8VLCAksith/view?usp=sharing" rel="noopener noreferrer" class="c-link"&gt;
            WhatsApp Video 2026-10-05 at 04.09.24.mp4 - Google Drive
          &lt;/a&gt;
        &lt;/h2&gt;
        &lt;div class="color-secondary fs-s flex items-center"&gt;
            &lt;img alt="favicon" class="c-embed__favicon m-0 mr-2 radius-0" src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fssl.gstatic.com%2Fdocs%2Fdoclist%2Fimages%2Fdrive_favicon_2026_32dp.png" width="32" height="32"&gt;
          drive.google.com
        &lt;/div&gt;
      &lt;/div&gt;
    &lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;A typical interaction looks like:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"2 roti, aadha katori rajma, thoda paneer aur ek glass chaas"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Instead of manually searching for each food, the AI interprets the meal, identifies the relevant foods and portions from NutriGhar's food database, and converts them into structured entries.&lt;/p&gt;

&lt;p&gt;The app can then update the user's daily calorie and protein totals and help them understand what they still need to eat.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/ArmanSinha7" rel="noopener noreferrer"&gt;
        ArmanSinha7
      &lt;/a&gt; / &lt;a href="https://github.com/ArmanSinha7/NutriGhar" rel="noopener noreferrer"&gt;
        NutriGhar
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      HacktoberFest work week 1 based on a app built for a friend theme helping to make a friend do well on their weight loss and muscle gain journey tracking their daily intake properly
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;NutriGhar 🥗&lt;/h1&gt;
&lt;/div&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;An offline-first Indian nutrition and weight-management companion built for my sister and best friend, Ayushi Sinha.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;NutriGhar started with a simple request from my sister, Ayushi: she wanted an easier way to track her daily calories and protein while working toward losing weight and getting fitter.&lt;/p&gt;
&lt;p&gt;Instead of building another generic calorie counter, I wanted to make something that understands the way we actually talk about food in India — from roti and dal to katori-sized portions and everyday Hinglish descriptions.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;What NutriGhar Does&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;NutriGhar is a Flutter-based Android application for personal nutrition tracking.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h3 class="heading-element"&gt;🥗 Food &amp;amp; Meal Tracking&lt;/h3&gt;
&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;Search a large Indian-focused food database&lt;/li&gt;
&lt;li&gt;Log meals and portions&lt;/li&gt;
&lt;li&gt;Support common Indian food names and aliases&lt;/li&gt;
&lt;li&gt;Track daily calories and protein&lt;/li&gt;
&lt;li&gt;Add custom foods&lt;/li&gt;
&lt;/ul&gt;
&lt;div class="markdown-heading"&gt;
&lt;h3 class="heading-element"&gt;🎯 Personal Goals&lt;/h3&gt;

&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;Set nutrition and weight goals&lt;/li&gt;
&lt;li&gt;Calculate calorie and protein targets&lt;/li&gt;
&lt;li&gt;Track progress toward personal goals&lt;/li&gt;
&lt;li&gt;Monitor weight trends&lt;/li&gt;
&lt;/ul&gt;
&lt;div class="markdown-heading"&gt;
&lt;h3 class="heading-element"&gt;📊 Progress &amp;amp;&lt;/h3&gt;…&lt;/div&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/ArmanSinha7/NutriGhar" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;NutriGhar is built with &lt;strong&gt;Flutter and Dart&lt;/strong&gt; and follows an offline-first approach.&lt;/p&gt;

&lt;p&gt;The project includes an Indian food database, food search, meal logging, portion handling, nutrition calculations, goals, progress tracking, and local persistence.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;The original NutriGhar application provided the core nutrition infrastructure: Indian food data, food search, portion handling, deterministic nutrition calculations, goals, meal logging, and progress tracking.&lt;/p&gt;

&lt;p&gt;The biggest weakness was the way users had to express what they ate.&lt;/p&gt;

&lt;p&gt;Traditional rule-based parsing works for clean inputs, but real people don't always describe food in perfectly structured terms.&lt;/p&gt;

&lt;p&gt;They say things like:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"2 roti aur thoda paneer"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;or&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"aadha katori dal aur ek glass chaas"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;and often mix English, Hindi, and Hinglish in the same sentence.&lt;/p&gt;

&lt;p&gt;So I focused the new AI work on that exact problem.&lt;/p&gt;

&lt;p&gt;I integrated &lt;strong&gt;Gemma 3 4B&lt;/strong&gt;, an open-weight model, running locally through &lt;strong&gt;Ollama&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;The AI layer takes a natural-language meal description and extracts structured information such as:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;food item&lt;/li&gt;
&lt;li&gt;quantity&lt;/li&gt;
&lt;li&gt;portion/unit&lt;/li&gt;
&lt;li&gt;additional notes&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The system then uses NutriGhar's food-search layer to retrieve relevant candidates from the local food database and grounds the model's output against those candidates.&lt;/p&gt;

&lt;p&gt;This is important because the AI is &lt;strong&gt;not responsible for inventing nutrition values or doing the arithmetic&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;The model understands the user's language.&lt;/p&gt;

&lt;p&gt;The application performs the nutrition calculations.&lt;/p&gt;

&lt;p&gt;The pipeline is:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Natural meal description → Gemma → structured extraction → food candidate retrieval → grounded food selection → deterministic nutrition calculation → daily totals&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;I also kept the AI layer replaceable so that different open models can be evaluated without rewriting the rest of the application.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;Nutrition data is personal.&lt;/p&gt;

&lt;p&gt;Meals, weight, fitness goals, and eating habits are information that someone may not want continuously sent to a third-party AI provider.&lt;/p&gt;

&lt;p&gt;That's why I wanted the AI component to work locally.&lt;/p&gt;

&lt;p&gt;With &lt;strong&gt;Gemma running through Ollama&lt;/strong&gt;, meal descriptions can be processed locally rather than being sent to a proprietary AI API. This provides a stronger privacy model and avoids a per-request API cost.&lt;/p&gt;

&lt;p&gt;Open models also give me more control over the system.&lt;/p&gt;

&lt;p&gt;I can experiment with different models, adapt the application to Indian and Hinglish food terminology, and potentially fine-tune a model for the specific way people describe their meals.&lt;/p&gt;

&lt;p&gt;This matters for phrases such as:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"aadha katori"&lt;/p&gt;

&lt;p&gt;"thoda paneer"&lt;/p&gt;

&lt;p&gt;"mess wali dal"&lt;/p&gt;

&lt;p&gt;"2 roti"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;A general-purpose nutrition application doesn't necessarily understand the context behind these expressions.&lt;/p&gt;

&lt;p&gt;An open model gives me the ability to improve that understanding without making the entire application dependent on a closed API.&lt;/p&gt;

&lt;p&gt;The philosophy behind the system is:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The AI understands the meal. The application does the math.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Best Use of Gemma&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;I am entering the &lt;strong&gt;Best Use of Gemma&lt;/strong&gt; category because Gemma 3 4B is used as the open-weight AI model at the core of NutriGhar's natural-language meal understanding pipeline.&lt;/p&gt;

&lt;h2&gt;
  
  
  Earlier Work &amp;amp; Attribution
&lt;/h2&gt;

&lt;p&gt;NutriGhar began as a personal project inspired by my sister and best friend, Ayushi, asking me for a simple way to track her daily calories and protein.&lt;/p&gt;

&lt;p&gt;For this challenge, I expanded that foundation with a new AI-driven meal-understanding workflow focused on making natural food logging easier and more practical.&lt;/p&gt;

&lt;p&gt;The existing nutrition infrastructure and food data provide the foundation, while the new work focuses on integrating open AI into the part of the application where language understanding creates the most value.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Ayushi Thought
&lt;/h2&gt;

&lt;p&gt;I built this for someone I actually care about, so the most important feedback is hers.&lt;/p&gt;

&lt;p&gt;The goal of NutriGhar was never just to make another calorie tracker.&lt;/p&gt;

&lt;p&gt;It was to take a small problem my sister had and turn it into something she could actually use.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>weekendchallenge</category>
      <category>hf26challenge</category>
    </item>
  </channel>
</rss>
