Mobile AI track
BridgeLearn is an offline-first adaptive study coach designed for Arm-powered phones and tablets. It explores a practical question: how much useful personalization can fit entirely on-device, without an account, a learner's name, or a reliable connection?
What it does
A learner completes a five-step lesson loop in Math, Science, or Reading. A tiny local policy reads five numeric signals—confidence, accuracy, response time, hint use, and recent progress—and chooses one of four next actions: Hint, Concept, Practice, or Reflect.
The interface shows the chosen action and policy confidence. It also offers an easy-read view, a take-home note, local-only progress, and an aggregate teacher view that does not require learner names.
How we built it
The application uses TypeScript and React and is delivered as an installable offline-capable web app. A service worker caches the experience so a previously opened lesson remains usable without a network connection.
The adaptive policy is a compact 5→32→4 neural network trained with a deterministic synthetic curriculum. We export both FP32 and symmetric INT8 artifacts. Browser inference uses typed arrays and a custom INT8 path, with a deterministic rules-based fallback if the model cannot load. Training, quantization, tests, and benchmarks are reproducible from the public repository.
Arm optimization
BridgeLearn targets Arm-powered mobile devices where download size, cache pressure, memory use, and energy efficiency matter.
- FP32 model: 1,296 bytes
- INT8 model: 324 bytes
- Size reduction: 75%
- Validation accuracy: 97.25% for both FP32 and INT8 on 1,200 held-out samples
- Arm64 browser benchmark: 120,000 INT8 decisions in 76.7 ms on an Apple M5 Pro using Chrome, about 0.639 microseconds per decision
The repository includes the benchmark and environment details. We deliberately do not claim that JavaScript INT8 arithmetic is faster than native floating point; this path is optimized for a small transferable artifact and dependable offline caching.
Challenges
The hardest part was preserving policy behavior after quantization while keeping the implementation understandable and reproducible. We also had to make offline use a real product behavior, which required state recovery, a safe fallback, and privacy-friendly local progress from the beginning.
Accomplishments
- A polished learning experience that continues to work offline
- A 324-byte adaptive policy with no validation-accuracy loss after quantization
- Reproducible model artifacts and benchmark evidence
- No required account, learner name, cloud inference call, or analytics service
- A design suited to shared and older Arm devices
What we learned
A small model can create meaningful personalization when its signals and actions connect to a clear learning loop. On constrained devices, product design and model design must be optimized together.
What's next
We would expand the curriculum, validate the policy with educators and learners, add multilingual lesson packs, explore a native WebAssembly/Arm SIMD path, and measure energy and latency across more low-cost Arm phones and tablets.
Log in or sign up for Devpost to join the conversation.