An Agentic AI Learning Companion for Autistic Children
An On-Device, Privacy-First Tablet Companion for U.S. Children on the Autism Spectrum, Ages 3-8, Now Live in Production Across Partner Clinics
Agentic AI tablet companion with edge-only, on-device inference
Built, piloted, and shipped to production across three partner clinics
EdTech / Autism Early Intervention
BrightPath moved from build to a clinical pilot to a live deployment now running across three partner clinics. The assessment mini-games, the VB-MAPP-aligned learner model, and the secure on-device pipeline are all in daily use with children, and therapists are working from the one-tap dashboards that surface skill-mastery probabilities after each session.
The clearest engagement result is on attention. The timely micro-celebrations built into the mini-games lifted average sustained attention from the 8-10 minute range toward 12 minutes per session, meaningful in a context where every additional minute of focused interaction is time the intervention actually lands. Combined with the auto-generated next-session playlists, that has cut the manual prep therapists used to carry between sessions.
Just as important is what did not change: privacy. Because every camera frame and voice sample is processed on the device and only anonymized skill-event logs ever sync out, the deployment holds the line on data protection even at production scale, backed by granular per-sensor consent, differential privacy on telemetry, and quarterly bias audits across demographics.
BrightPath is a tablet app powered by agentic AI, an always-on partner that observes, learns, plans, and acts in the moment for each child. Built for U.S. children on the autism spectrum, ages three to eight, it screens cognitive, language, and socio-emotional skills in minutes, remixes lesson paths on the fly, and gauges engagement through private, on-device analysis.
The product is now live in production. After completing the MVP build and a clinical pilot, BrightPath is running across three partner clinics, with its assessment mini-games, learner model, and secure on-device pipeline all in daily use. Every bit of computer-vision and speech inference happens on the tablet itself, so the most sensitive data a family owns never leaves the device.
The Challenge
Families and early-intervention centers consistently report three pain points that BrightPath was built to address: milestone tracking that is inconsistent enough to cause late or missed interventions, a heavy therapist prep load, and short attention spans that cut sessions off before they do their work. The hard part was not any single feature, it was doing all of this on a children's device without sending sensitive video or audio off the tablet, so privacy was a design constraint from the first line of code rather than a compliance afterthought.
Inconsistent milestone tracking across sessions, which led to late or missed early interventions
High therapist prep load, roughly 20 minutes of preparation per child per session
Attention drift in longer sessions, with average interaction capping around 8 to 10 minutes
Sensitive inputs (camera frames, speech) that cannot be sent to the cloud, all inference has to run on the device
Wide ability ranges among children, so a one-size lesson path fails most of them
Older, school-issued tablets with limited hardware that still have to run computer-vision inference smoothly
The Solution
BrightPath runs as an agentic loop, Observe, Orient, Plan, Act, Learn, that operates continuously for each child, with privacy engineered into every phase. All computer-vision and speech inference happens fully on-device; only anonymized, non-identifying skill-event logs sync to Azure Government Cloud. Every element below is live in the shipped product.
Observe, on-device sensing
Touch, speech, and low-resolution camera frames are captured and processed locally on the tablet, so no raw photos or audio ever leave the device
Orient, Bayesian learner model
A learner model aligned to the VB-MAPP assessment framework estimates each child's skill-mastery probabilities and updates them as the session unfolds
Plan, reinforcement-learning policy
A Double DQN lesson-selection policy remixes the next activities to match each child's ability in real time
Act, multi-sensory mini-games
Assessment and practice mini-games for letters, counting, and emotion recognition, with timely micro-celebrations that lift sustained attention from roughly 8-10 minutes toward 12
Learn, summaries and red-flags
A daily summary generator and red-flag detection layer surface skill-mastery probabilities to therapists and auto-generate next-session playlists
Privacy and ethics by design
Edge-only image processing, granular per-sensor consent toggles, differential privacy on any telemetry export, and quarterly bias audits to verify parity across demographics
Why It Worked
BrightPath works in production because its hardest constraints, privacy and clinical credibility, were treated as features rather than obstacles. Processing every camera frame and voice sample on the device means the most sensitive data a family owns never travels, and building the agentic loop around a real assessment framework (VB-MAPP) rather than a generic chatbot means the product is accountable to clinical reality, not just to a demo.
BrightPath shows how Luzran builds agentic AI for a high-stakes, privacy-critical domain and takes it all the way to production: on-device inference so sensitive data never leaves the tablet, a learner model anchored to a recognized clinical framework, and a lesson policy that adapts to each child in real time, now live across partner clinics. If you are building an AI product where privacy, safety, and credibility are non-negotiable, this is the standard of engineering we hold ourselves to.
Your business could be next.
Every transformation starts with a conversation. Let's discuss what's possible.
No sales pitch. No guilt trip. Just clarity on what's costing you.