AIdeas Finalist: Ivy - The World's First Offline-Capable, Proactive AI Tutoring Agent
A proactive AI tutoring agent powered by Amazon Bedrock that monitors student progress, initiates outbound voice calls, and runs on-device inference for offline learning addressing educational inequity for millions of Ethiopian students with intermittent connectivity.
🎥 Watch the 3-minute demo first: https://www.youtube.com/watch?v=SK9-iTqDwH0
App Category: Social Impact
Tags: #aideas-2025 #aideas-2025-finalist #social-impact #EMEA
My Vision
Two months ago, I became a semi-finalist in the 10,000 AIdeas competition. I described a proactive AI tutoring system for Ethiopian students: one that initiates outbound voice calls when students go silent, renders interactive learning components inside the conversation, and runs a quantized language model on $80 phones without internet.
Since then, three things happened that validated the core thesis faster than I expected. PrismML released Bonsai, true 1-bit language models (8B parameters in 1.15 GB). Google released Gemma 4 with on-device audio and vision. And a team published Ethio-ASR, state-of-the-art Amharic speech recognition at 22.9% WER. Every one of these developments landed in the exact gap Ivy was engineered to fill.
Ivy is an AI tutoring agent built for 35 million Ethiopian students. In Ethiopia, the student-to-teacher ratio reaches 1:70 in rural areas. Private tutoring costs more than most families earn. Sixty percent of the population lacks reliable internet. National exams at Grades 8, 10, and 12 determine whether a student's education continues or ends.
My vision: every Ethiopian student gets a patient, bilingual, always-available AI tutor that speaks Amharic, understands their curriculum, works with or without internet, and refuses to give up on them when they stop showing up.
Why This Matters
Most AI education tools are reactive: you open the app, ask a question, get an answer. Then you navigate to a separate screen for a quiz. Another for flashcards. Another for your study plan. The student is constantly switching contexts, losing momentum.
More importantly, every AI education tool I've surveyed (Khanmigo, Duolingo Max, Quizlet Q-Chat, Socratic by Google) becomes completely useless when the internet drops. For a student in rural Tigray where connectivity is intermittent for hours, that means the tutor disappears precisely when it's needed most.
And none of them chase students down. They all wait passively for the student to show up. A 14-year-old preparing for a national exam alone doesn't need a tool that waits. She needs a tutor that cares enough to call.
Ivy addresses all three failures:
- Interactive Immersive Agentic Learning: The AI renders quizzes, flashcards, study plans, and explanations as interactive widgets directly inside the conversation. The chat IS the classroom. Zero navigation. Zero context switching.
- Proactive Autonomous Intervention: A background engine monitors study schedules, overdue spaced repetition cards, and exam proximity. When a student goes silent, Ivy escalates from nudge to notification to outbound voice call, with a personalized script generated in real-time by Amazon Bedrock.
- Offline-first architecture with on-device AI: Three layers of offline capability: pre-cached study packs, on-device inference via the Ivy Model family, and on-device Amharic speech recognition via Ivy Voice.
How I Built This
Ivy runs on a deep AWS-native architecture spanning 8 services.
The Cognitive Core: Amazon Bedrock runs Claude Sonnet 4 with 21 custom AI tools in ReAct reasoning mode. The agent autonomously chains tools (checking progress, searching course materials, generating quizzes, creating study plans) and synthesizes results into interactive UI components that render inline.
The Memory Layer: Five Amazon DynamoDB tables form a student knowledge graph: learning events, conversational memory, curriculum data, study schedules, and a proactive nudge queue. Single-digit-millisecond latency.
The Proactive Engine: A 5-minute cron process scans all active students across three dimensions: missed sessions, overdue SRS flashcards, and exam proximity. It escalates through in-app nudges, push notifications, and outbound voice calls. The voice calls aren't pre-recorded. Amazon Bedrock generates personalized scripts based on each student's specific situation.
The Voice Pipeline: Amazon Nova Sonic (
amazon.nova-sonic-v1:0) as the primary voice engine, providing unified speech-to-speech via Bedrock's bidirectional streaming API with tool use support and natural turn-taking. Gemini serves as fallback for Amharic until Nova Sonic adds language support. Outbound call scripts are generated by Claude Sonnet 4, then synthesized through Nova Sonic.The Edge Module: A native Android application in Kotlin (MVVM, StateFlow, Coroutines) that runs on-device inference with automatic model discovery, GPU acceleration, and Ivy's Socratic system prompt baked in, both English and Amharic.
Deployment: Docker containers on AWS App Runner (auto-scaling), Amazon ECR for versioned images, Amazon S3 for audio and course materials, AWS IAM for least-privilege access, Amazon CloudWatch for observability. Total AWS infrastructure cost at current development scale: under $200. At scale, the economics are fundamentally different from cloud-only competitors: every student running on-device inference costs Ivy exactly $0 per query. The server handles only proactive orchestration, course ingestion, and memory sync, not inference.
Agentic IDE and CLI: Kiro To accelerate the development of Ivy’s 21-tool agentic framework, I used the Kiro agentic IDE and CLI. Kiro was instrumental in architecting the complex logic between the Claude-powered reasoning engine and the local quantized inference on-device.
Projected at-scale costs (AWS on-demand pricing):
| Students | App Runner (1 vCPU/2GB) | DynamoDB (on-demand) | S3 + ECR | Bedrock (proactive scripts only) | Total/month |
|---|---|---|---|---|---|
| 1,000 | ~$7/month (1 instance) | ~$3/month | ~$2/month | ~$5/month | ~$17 |
| 10,000 | ~$7/month (1 instance) | ~$15/month | ~$5/month | ~$15/month | ~$42 |
| 100,000 | ~$14/month (2 instances) | ~$80/month | ~$15/month | ~$50/month | ~$159 |
The key insight: on-device inference doesn't replace AWS, it extends it. The standard version runs fully on Amazon Bedrock. Ivy Edge adds a third layer for students with no connectivity at all. For Ivy Edge users, the server-side cost drops to $0.0016/student/month at 100,000 students, since the server handles only proactive orchestration, memory sync, and course ingestion. The architecture makes AWS-powered education economically viable at national scale in a market where cloud-only would be unaffordable.
Privacy & Student Data Protection: GDPR-compliant consent flow with granular toggles (data processing, AI tutoring, voice, proactive outreach, third-party AI). Parental consent required for students under 16 (GDPR Article 8 / COPPA). Four data subject rights as live API endpoints: Right of Access, Data Portability (JSON export), Erasure (cascading deletion across all five DynamoDB tables), and Right to Object. AES-256 encryption at rest, Helmet security headers, tiered rate limiting, FERPA-compliant access controls. On-device via Ivy Edge: student data never leaves the phone.
The Client: React Native + Expo with 5-tab interface, AI SDK v4 streaming, full Amharic localization (250+ strings), animated mascot, gamification (XP, 10 Ethiopian-themed levels, 9 achievement badges), and SM-2 spaced repetition that works offline.
Demo
What I Learned
Which judge feedback resonated most
The judges identified five areas for refinement. Two hit hardest: "Roadmap vs. Reality" and "Limited Validation Data."
On security: they were right that my original submission said nothing about student data protection. I rebuilt the entire security layer, and the Privacy section above describes what now exists in the codebase.
On roadmap: IoT Greengrass edge nodes remain a roadmap item with zero code. Ethio-ASR has an architecturally prepared pipeline but the native bridge is not yet connected. Everything else in "How I Built This" is implemented and running. I'd rather be honest about the gap than blur it.
Since submission, the ecosystem moved: Bonsai 1-bit models (PrismML) gave Ivy true 1-bit quantization. I integrated three models as the Ivy Model lineup, with the 1.7B model delivering 70% more parameters than my original Gemma 3n 1B in 55% less space. Gemma 4 E4B added on-device vision. Ethio-ASR achieved 22.9% WER on Amharic.
But the validation feedback hit hardest, because they were right. So I ran a pilot.
Phase 1 Pilot: 10 Students, Addis Ababa
Over one week, 10 high school students (Grades 9–12) across Addis Ababa used Ivy as their primary study companion across five subjects: biology, chemistry, physics, history, and mathematics. This wasn't a controlled experiment with a data science protocol. It was a field test: real students, real phones, real study pressure.
Proactive Autonomous Intervention: All 10 students received proactive calls from Ivy. All 10 picked up. The reactions were a mix. Some were genuinely shocked that an AI was calling them, others laughed before engaging, but every single student who accepted a call went directly into a study session. 100% call acceptance. 100% immediate engagement. No student ignored Ivy when she reached out. For context, typical push notification open rates in education apps hover between 5–15%. Ivy called, and every student answered.
Interactive Immersive Agentic Learning: Students described the inline interactive components (quizzes appearing mid-conversation, flashcards generated from the lesson they just learned, study plans built around their weak areas) as feeling like "magic." The continuity was the key insight: a student could ask Ivy to explain cellular respiration, get tutored through it Socratically, take a quiz on it, practice flashcards for retention, and test their understanding, all in the same scrollable conversation. No screen changes. No navigation. No momentum lost. They didn't want to stop.
Three-Layer Offline Capability: Students tested Ivy Edge on Samsung A14 phones ($80 devices here in Ethiopia). The on-device model performed well, and students used it continuously to study without a single network connection. But the three-layer architecture proved its value most for students with partial connectivity, intermittent 2G or 3G. These students used Ivy's standard cloud mode when the connection held, then seamlessly fell back to pre-cached quizzes and flashcards when it didn't, without needing to switch to the full Edge mode. The layered design meant learning never stopped regardless of connectivity state.
After the week, all 10 students asked to keep access. 100% retention.
This is a small sample. It's not a peer-reviewed study. But it's real students on real phones in the city where I built this, and every core innovation worked as designed.
External Technical Validation
My project title AI won 1st Place in the Amazon Nova Hackathon UI Automation catagory, a $40,000 prize pool competition judged by AWS VPs and Jeff Barr. Different domain (Nova Act-powered web automation), but the same architecture pattern: autonomous agents with multi-tool orchestration on AWS. The same ReAct reasoning and agentic design that powers Ivy's 21-tool tutoring system.
Separately, Violette Naa Adoley Allotey, founder of SchoolOS in Ghana (a top-300 AIdeas submission), reached out requesting integration of Ivy's offline AI into her school management platform. Unsolicited. Ivy's offline architecture isn't just solving an Ethiopian problem.
On competitive landscape
The judges asked why Khanmigo and Duolingo Max haven't implemented proactive intervention or offline-first architecture. Three structural reasons: business model misalignment (cloud-first SaaS products can't push inference to devices without eliminating their revenue mechanism), market focus (three-layer offline has no ROI in markets with reliable connectivity), and uncomfortable design choices (an AI that calls you when you skip studying is right for a student facing a life-determining exam, but it would be perceived as invasive in casual language learning).
To be concrete:
| Capability | Khanmigo | Duolingo Max | Quizlet Q-Chat | Ivy |
|---|---|---|---|---|
| Proactive outbound calls | No | No | No | Yes |
| Inline interactive components (IIAL) | No | No | No | Yes |
| On-device AI inference (offline) | No | No | No | Yes |
| Pre-cached offline study packs | No | Limited | No | Yes |
| Amharic / Ge'ez script support | No | No | No | Yes |
| Spaced repetition (SM-2) | No | Proprietary | Basic | SM-2 |
| Free / zero marginal cost | $44/yr | $84/yr | $36/yr | $0 |
Every competitor charges a subscription. Every competitor requires internet. None of them call students back.
Key development insights
Amharic's Ge'ez script (ፊደል) is underrepresented in most LLMs. Testing revealed that Claude via Bedrock handles conversational Amharic, Socratic dialogue, and quiz generation with quality sufficient for genuine learning. Not perfect. Good enough. That distinction matters more than any benchmark.
1-bit quantization isn't just compression, it's a new deployment paradigm. A 250 MB download gives a student an AI tutor for life. When hardware designed for 1-bit inference arrives, these models will be another order of magnitude faster. The economics of education just changed.
Intelligence must be local, affordable, and linguistically native. Ivy was architectured for this moment. The pieces arrived faster than anyone expected.
Ivy grows everywhere.
Built by a solo developer from Addis Ababa, Ethiopia.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article