BlockScapes

What it does

Kids snap together alphabet blocks under a camera; YOLOv8 reads the letters, MarianMT checks the word against a Grade-1 list, and the kit erupts in green cheers or gentle red hints—turning spelling drills into fast, feedback-rich play.

Your inspiration

Classroom logs in our project document show Grade 1 pupils spending whole sessions on copy-the-board drills yet receiving no correction until papers are marked hours later—so the same misspellings resurface the next day. Our tests of the three leading e-toys cited in the report found plenty of flashing lights but no meaningful guidance, and tablet apps, while interactive, stripped away the motor-memory benefits of real blocks. This clear gap—delayed feedback plus shallow digital engagement—shaped our brief: fuse tactile play with an on-device AI checker that gives every child instant correction, upbeat reinforcement and trackable progress.

How it works

A Raspberry Pi 4, fan-cooled within a 3D-printed shell, anchors a camera ringed by dimmable LEDs. After a spoken prompt, the child assembles a word inside the lit frame. YOLOv8, fine-tuned on our block images, detects each letter and concatenates them. MarianMT, chosen for its speed-accuracy balance, checks the attempt against a curated 2–5-letter Grade-1 list. A correct word triggers green flashes and cheers; an error gets red light and a phonetic hint. Attempts, timings and safe IDs flow through Express APIs to a MongoDB store, while a React dashboard charts streaks for teachers or parents. All compute stays local, so no internet is required and pupil data remain private.

Design process

Interviews with teachers, pupils and a child-development expert confirmed the need for tactile play plus instant feedback. Benchmarking LeapReader, Osmo and ABCmouse set five constraints: runtime, maintainability, training cost, latency and F1-score. A cardboard rig with printed letters verified camera height and angles. MarianMT, DistilBERT and ELECTRA were trained on a 4,458-word corpus; Pareto analysis crowned MarianMT for accuracy-per-watt. YOLOv8, labelled over four cycles, pushed detection past 98%. A 3D-printed case locked the Raspberry Pi, camera, LEDs and buttons; three iterations cut glare and refined vents. Python vision/NLP and a MERN logger were integrated, and seven test sets × 20 trials achieved 99.9% accuracy. Classroom pilots then damped fan noise and tuned colors for WCAG 2.2, yielding the current kit.

How it is different

Most spelling toys merely flash lights, and app-only tools sacrifice the tactile element. Our system fuses real blocks with edge AI: YOLOv8 recognizes any orientation without QR markers, MarianMT runs fully offline on a Raspberry Pi, freeing schools from subscriptions and protecting data, while an open-source MERN dashboard visualizes progress. The hardware meets ASTM F963, ISO 8124 and WCAG 2.2, setting it apart from hobbyist builds and commercial kits alike.

Future plans

We will use 8-bit-quantise MarianMT and migrate to Raspberry Pi 5 with NPU support to cut latency, publish our block-detection dataset and open-source the bill of materials. An educator portal will let teachers upload wordlists and regional languages. A pilot batch of classroom kits and workshops will precede a longitudinal literacy study. Longer term, we aim to license the design to toy makers and add Bluetooth blocks that auto-sense letter orientation for richer games.

Awards