How I Took Offline Textbook Search from 73% to 87% Accuracy on Android
Building AI answers from NCERT textbooks with no internet on a classroom panel. The model was the easy part; the PDFs, the chunking and the size of the index decided accuracy.
Shubham Singh BishtAndroid Systems & Edge AI Engineer
Shubham Singh Bisht: I build Android systems and explore AI at the edge.
I'm Shubham Singh Bisht, an Android engineer based in Bengaluru. Currently building high-performance system software and interactive displays at Teachmint (ex-Pratilipi), with a focus on runtime internals, low-latency graphics, and on-device AI.
9:41
9:41
Today
Android software & system layers for interactive flat-panel (IFP) devices
Work
currently.kt
- Building
- Android software & system layers for interactive flat-panel (IFP) devices
- Exploring
- On-device LLM/SLM inference, NPU acceleration, and local RAG pipelines
- Investigating
- Android runtime (ART), coroutine schedulers, and IPC boundaries
Projects
Writing
Terminal
About
Settings
Wallpaper & style
Display
Notifications
Developer options
About phone
ModelPortfolio 1 · Bengaluru
Android version16
Whiteboard
Draw here ✎
How it works
Android apps on 4K classroom panels
Teachmint's classroom panels are big Android screens on the classroom wall. I work on the whiteboard and classroom apps that run on them, where recording, startup, pens and firmware behave nothing like a phone.
- A recorder that keeps lessons safe
- Faster start-up, measured across the fleet
- Works with pens and firmware from several vendors
- Fewer lesson recordings lost per panel to crashes and kills, new recorder vs old
- 94%
- Time for the whiteboard app to open, against the production build, measured on a panel (~25% faster)
- 2,430 ms → 1,807 ms
- Time to save a two-hour lesson after pressing stop
- ~4 min → 2.4 s
AI Lab
Explain photosynthesis in one line.
Plants turn light, water and CO₂ into glucose and oxygen.
Demo— tok/soffline
How it works
Language, vision and speech AI with no internet
A classroom panel has three chips that can run AI: the processor, the graphics chip and an AI accelerator. I measured which one each model should use, and moved AI into its own process so a crash can't take the whiteboard down with it.
- Runs language, image and speech models with no internet
- Picks the right chip for each model
- Keeps AI in its own process, so a crash can't take the whiteboard down
- Faster on the panel's graphics chip than its processor (3.07 vs 0.45 tokens per second)
- ~7×
- Time for the AI model to start, after switching to a smaller model
- 38 s → 5 s
- Wait for speech to appear as text, after fixing a threading bug
- 14–52 s → 0.3–1.2 s
Recorder
How it works
Lesson recordings that survive crashes
Teachers record whole lessons on the panel, and often that recording is the lesson's only copy. The old recorder lost a file whenever the app crashed or was swiped away. The new one records in five-minute pieces that are each playable, then joins them when the lesson ends.
- Records in five-minute pieces, each one playable
- Survives crashes, swipe-kills and other apps
- Tells the team when a recording fails
- Fewer recordings lost per panel to crashes and kills, new engine vs old, over the same four weeks
- 94%
- Time to save a long lesson after pressing stop (2-hour test → a real 113-minute lecture)
- ~4 min → 2.4 s
- Longest stretch a crash can cost, because every five-minute piece is a playable file
- 5 min
Textbooks
How do plants make food?
- Biology · ch.7 · p.112 · 0.92Photosynthesis takes place in the chloroplast…
- Biology · ch.7 · p.114 · 0.81Chlorophyll absorbs light energy…
- Science · ch.2 · p.31 · 0.64Leaves exchange gases through stomata…
3 passages · 640 tokens → on-device model
How it works
Answers from NCERT textbooks, offline
An AI assistant is only as good as the passage it reads. This pipeline turns NCERT textbook PDFs into a database the panel can search with no internet, and it searches only the student's own class and subject.
- Turns textbook PDFs into a searchable database
- Searches only the student's own class and subject
- Catches bad PDFs, like draft chapters and broken Hindi fonts
- Questions where the right textbook passage came first, once search was limited to the student's class and subject
- 73% → 87%
- Textbook passages made searchable, NCERT Classes 1–12 in English
- 32,338
- Typical search time on the panel, with no internet
- ~7–8 ms
Edu AI
How it works
A classroom AI assistant in 12 languages
Edu AI is an assistant that floats over whatever a teacher is doing on a classroom panel: select anything on screen and ask about it. I led taking it beyond India, to new panel brands, new languages and right-to-left layouts.
- Floats over any app on a classroom panel
- Select anything on screen and ask about it
- Prompt changes without an app update
- Vendor classes in our own panels' apps, checked in the built APKs
- 0
- Languages shipped, including Arabic and Persian
- 12
- Panel brands supported from one codebase
- 4
Sign-In
Whiteboard · signed in as Priya
Choose an account
Shared across Whiteboard and EduAI
How it works
One login, shared safely between two apps
A teacher signs in to the whiteboard once. The AI assistant on the same panel asks the whiteboard who is signed in, proves it's a trusted app, and gets a one-time code. It never sees the teacher's password or session.
- Checks which app is asking before it answers
- Hands over a one-time code, never a password
- Reuses the whiteboard's own login screen
- Passwords or tokens passed between the two apps
- 0
- How long the one-time sign-in code lasts
- 60 s
- Differently signed versions of the assistant it has to trust
- 4
Daybook
How it works
A private journal of my workday
A Mac app I built for myself. It listens through the day, transcribes Hindi–English conversation, works out who said what, and writes a daily report against my goals. Nothing leaves the laptop.
- Transcribes Hindi–English conversation
- Knows who said what
- Writes a daily report against my goals
- Words captured in 12 real recordings, after fixing speech-to-text on long audio
- 65% → 98.5%
- Conversations counted in one day, once phone and Mac recordings of the same room were merged
- 235 → 62
- Time to write the daily report on an AI model running on the laptop
- 9 min → 4 min
Updater
How it works
Panels that update themselves
Classroom panels have nobody standing next to them to tap Update. This SDK lets a preinstalled app download, resume and install its own updates on four brands of panel, including one whose firmware refused the permission it needed.
- Downloads, checks and installs its own updates
- Picks up where it left off after a crash or reboot
- Works on firmware that refused the permission it needed
- Panel brands updated from one SDK
- 4
- App updates that pick up where they left off after a crash or reboot
- 100–180 MB
- On one brand's first update, then automatic (before: every update failed)
- 1 tap
EduBits
How it works
Web products inside the Android app
Teachmint's AI tutor and EduBits are web apps shown inside the native Android app. I turned one-off WebViews into a shared platform: a bridge that waits until the page is ready, and web screens that stay alive as you move around.
- A bridge that waits until the web page is ready
- Web screens that stay alive as you move around
- Clear errors instead of blank screens
- Web screens kept alive between pages, instead of reloading on every screen
- 3
- Named error types, so a blank screen tells us why
- 5
League
How it works
Leaderboards, streaks and a social feed
Three features to give students a reason to come back: a weekly leaderboard, Boost with XP and streaks, and a social feed. Likes, updates and sharing all had to feel instant on low-end phones.
- A weekly leaderboard league
- Boost, with XP and streaks
- A social feed with likes and sharing
- Student-facing features shipped in three months
- 3
- How often the feed checks for updates, set by the server and switchable off without an app update
- 30 s – 10 min
9:41Today
Notifications
- 10
- 15
Selected work
A few things I've built, investigated, and learned from.
How Coroutine Dispatchers Actually Work
Dispatchers.IO is not a pool of IO threads. What the CoroutineScheduler really does, and the mental model that replaced mine.
What Actually Happens When a Suspend Function Makes a Network Call?
If the thread is released while a request is in flight, what is actually waiting for the response? Coroutines, threads, sockets, and continuations, separated.
03