Om Gholwe — AI that runs where the internet doesn't

On-device AI · Flutter · Gemma 4 E2B

Learnify

A study companion that generates every lesson locally. One ~2.58 GB Gemma 4 E2B model runs through LiteRT-LM on the phone, and eight system prompts turn it into eight specialists. After the one-time model download there are no accounts, no analytics and no network calls — airplane mode works perfectly.

FlutterGemma 4 E2B LiteRT-LMSQLite MultimodalOffline-first

Repo ↗

📖 StoryVisual-novel lessons taught by 2–4 characters

🧠 TutorPlain-language explanations at your level

❓ QuizQuestions auto-targeted at your past mistakes

🗺️ MasterySplits a topic into 5–7 progressive steps

🔍 ExplorerBreaks topics into bite-sized sub-topics

📅 Planner7-day study plans from your real history

👤 Learner TwinChat that remembers every prior session

📷 VisionReads a textbook photo, extracts the topic

🗺️

Mastery Path

The spine of the app. Any topic becomes 5–7 gated steps — pass a quiz at ≥70% and the next step unlocks. Duolingo's loop, for literally anything.

🧠

Memory that compounds

Every quiz, wrong answer and chat turn is written to SQLite. The last 8 turns are replayed into each prompt, so the tutor opens knowing you missed ATP synthesis twice.

😊

Mood-aware voice

A daily check-in — calm, hyped, curious, anxious, low — is threaded silently into every prompt. The model never states your mood; the lesson just feels like it gets you.

🧑‍🏫

Feynman Mode

Unlocks at mastery, then flips the roles: the character becomes your confused junior and you teach it back. The best-evidenced study technique, wrapped in something a kid wants to do.

🎬

Franchise Lab

80 franchises × 6 characters, each with its own speech and humour style, so concepts arrive in a voice the learner already loves. Mood routing surfaces the right cast.

🖼️

Comic Album

Any lesson saves as a 4-panel manga grid, rendered in pure Flutter — no image generation, instant. Weeks of study become an artifact you can print.

♿

Accessibility skin

Opt-in dyslexia mode (Atkinson Hyperlegible, bionic reading, simplified prompts) and TTS with live karaoke word highlighting. Zero footprint when off.

🧬

Knowledge graph

A concept map and skill tree built from your real database rather than a fixed curriculum — they start empty and fill in as you learn.

Why on-device matters. Weak rural signal, blocked campus Wi-Fi, 2G, long commutes, power cuts. Cloud tutors fail exactly where tutoring is scarcest — so the model ships with the app and never phones home.

Safety systems · Flutter · Firebase

Raksha

A women's safety app built around one uncomfortable assumption: in a real emergency you cannot unlock your phone, open an app and press a button. So Raksha offers four independent ways to raise an alert — including one that is invisible to anyone watching the screen.

FlutterFirebase Google MapsSensors Background serviceFCM

Private repo

📳

Shake to alert

A background service watches the accelerometer and fires when three shakes land inside one second — no unlock, no screen, phone still in your bag.

🔊

Scream detection

Continuous microphone monitoring triggers an SOS when sound holds above 85 dB for two seconds, covering the case where the phone is out of reach entirely.

🧮

Calculator disguise

The app can hide as a working calculator. Five rapid volume-button presses fire a silent SOS; seven taps on the display restore the real interface.

📍

Priority contact alerts

On activation it resolves your location, writes an alert record, and SMSes each emergency contact in priority order with a live location link and your custom message.

🗺️

Community risk map

User-reported incidents cluster into ~200 m grid cells and render as a colour-coded heat map — green, orange, red — instead of sitting in a database unused.

🧭

Safe route scoring

Routes are scored 0–100% against known risk zones, with warnings on high-risk segments and nearby police stations and hospitals surfaced along the way.

🚨

Proximity warnings

Real-time monitoring pushes a notification when an incident is reported inside your configured radius, so community reports become early warnings.

🧳

Journey & safe zones

Trip tracking with live location sharing, trusted safe-zone geofencing, a fake-call escape hatch, and daily contextual safety tips.

The design bet. Motion detection false-fires; audio detection drowns in noise; a visible button can provoke an attacker. Each trigger covers another's failure mode — that redundancy is the product.

Team project · React · Node.js · MongoDB

AI-Based Software Development Platform

A collaborative workspace that folds AI assistance into the place work actually happens. Most tools solve either project management or code collaboration; this one puts both behind a single AI-assisted surface so a team is not stitching three products together.

ReactNode.js MongoDBReal-time

👥

Projects & collaborators

Create a project, add collaborators, and keep ownership and membership in one place rather than spread across separate tools.

💬

Real-time communication

Team messaging lives inside the development environment, so discussion stays attached to the work it is about.

🤖

AI assistance

Integrated help for coding and development tasks, aimed at the routine work that slows a team down between the interesting problems.

📊

One coordination surface

Management and collaboration in a single platform, reducing the context switching that costs distributed teams their momentum.

AI that runs where the internet doesn't.