Multimodal Conversation
Ask questions alongside images, photos, or PDF documents. The vision model reads and reasons about what it sees — attachments become part of the conversation.
Ask Harold
A private algorithmic assistant built for the questions you actually ask.
Harold is an AI-powered assistant built with privacy and practicality at its core. Available on iOS and Android, it brings the capabilities of frontier language models — including multimodal vision understanding — to a focused, thoughtfully designed mobile experience that doesn't compromise on data principles or user dignity.
Built on a Flutter foundation with a Cloudflare Workers backend, Harold processes requests at the network edge rather than in centralized server farms, keeping inference fast, architecturally lean, and under careful governance. Conversation history is tied to your account, exportable as a PDF, and deletable in full at any time.
Harold is not a general-purpose chatbot. It is designed around a specific personality, a clear purpose, and appropriate limits. It anchors a growing family of specialized assistants built on the same technical infrastructure, including Ask Tutor, a COPPA-compliant learning companion for students. Both emerge from a single shared codebase.
Architecture
Edge inference, thoughtful design
Harold runs on a Cloudflare Workers backend routing requests to AI models on Cloudflare's global edge network. Two models handle different kinds of requests: a compact 3-billion-parameter text model for everyday queries, and an 11-billion-parameter vision model when images, photos, or documents enter the conversation.
Before any response is returned, a dedicated safety model reviews AI output against thirteen harm categories. Crucially, moderation is applied to AI responses only, not user input. This deliberate choice eliminates the false positives that frustrate users of most content-filtering implementations.
Both Harold and Ask Tutor are built from a single Flutter codebase using a flavor system. Platform identity, branding, AI personality, feature flags, and compliance requirements — including COPPA compliance for the tutor — are all determined by a single build-time parameter. One codebase, two distinct products.
Technical Stack
- Flutter (Dart) Cross-platform iOS + Android
- Cloudflare Workers Serverless edge backend
- Cloudflare D1 Persistent conversation history
- Llama 3.2 (3B + 11B) Text and vision inference
- Llama Guard 3 AI output safety filtering
- RevenueCat Subscription management
- Google Sign-In / Sign in with Apple Authentication
Capabilities
What Harold does
Privacy-First by Design
Conversations are stored under your account, not sold or shared. Full data deletion is available on demand — no dark patterns, no buried settings.
PDF Export
Export any conversation to a clean, formatted PDF, ideal for saving research sessions, technical walkthroughs, or anything worth keeping.
Streaming Responses
Responses arrive as a live stream via Server-Sent Events — no waiting for the full answer to generate before reading begins.
Biometric App Lock
Conversations are protected by device biometrics — Face ID, Touch ID, or fingerprint — using secure hardware-backed storage.
Custom Instructions
Inject your own persistent system instructions to tailor Harold's tone, focus, and behavior to exactly how you like to work.
What's Next
Harold is part of something larger.
Ask Harold and Ask Tutor are the first in a planned family of purpose-built assistants — each with its own personality, domain, and constraints.