AutiHD rebuilds Mindory’s production AI with Amazon Bedrock AgentCore

The challenge: AI architecture was slowing Mindory’s roadmap

AutiHD is the Luxembourg company behind Mindory, an AI planner and companion designed around the needs of people with ADHD and autism. The company had already built and launched its first AI experience on AWS using Amazon Bedrock and multiple Anthropic models.

The initial system supported model routing, task creation, stress-related interactions and integration with the Mindory mobile application. As the product grew, the AI needed to remember users across sessions, perform structured actions reliably and follow more detailed behavioural guidance.

Each interaction rebuilt a static system prompt of roughly 9,000 tokens. The application sent that prompt alongside the last 20 messages, user information, task data and formatting requirements. Structured JSON output depended on prompt instructions and could truncate when the token budget ran out, breaking task creation.

Important context from older conversations disappeared at the daily reset. Voice transcription used an English-only batch process with a polling loop that timed out at roughly 30 seconds. Each new behaviour made an already large prompt harder to maintain.

AutiHD had capable developers. It needed specialist experience in AWS AI services and production AI architecture to move its roadmap forward as it prepared for its next stage of growth and a new funding round.

AWS referred AutiHD to CloudHero

AutiHD approached AWS with questions about architecture, performance and the cost of scaling its AI stack. AWS referred the company to CloudHero based on experience with similar AI workloads and demonstrated AWS competencies.

Technical discovery showed that the first priority was deciding which architecture AutiHD wanted to scale. A cost forecast for the existing implementation would have limited value if the team needed to replace its foundations.

AutiHD wanted faster, more intelligent and more personalised interactions, supported by a separate AI service behind its existing product. The engagement developed into a production AI engineering project.

From a monolithic prompt to production AI on Amazon Bedrock AgentCore

CloudHero designed the new AI layer as a separate application around Amazon Bedrock and Amazon Bedrock AgentCore. Incoming interactions are classified by intent and context, then routed to the appropriate execution path.

Separate components handle conversational support, tasks and higher-risk situations. Predictable calculations remain in deterministic software, while models handle reasoning and natural-language interaction.

Persistent memory for individual users

Amazon Bedrock AgentCore Memory retains relevant preferences, routines, ongoing stressors, communication preferences and techniques that previously helped a user. Context can remain useful across conversations without replaying large amounts of chat history.

The extraction policy stores short facts from user messages only. It excludes assistant turns and diagnoses the user has not stated. Memory remains associated with the individual user and can be retrieved, updated and deleted, giving AutiHD a foundation for user-facing memory controls.

Expert guidance expressed as structured rules

AutiHD supplied an expert-authored behavioural framework that grew beyond 100 pages. It described situations, response sequences, branching logic, personalisation and explicit restrictions.

CloudHero translated that framework into machine-readable rules. At handover, the system included 26 defined situations. A mandatory safety check runs first and overrides other routing. A classifier on Claude Haiku 4.5 then selects the relevant situation.

The selected execution path receives approximately 2,600 tokens of relevant guidance. Conversational support and crisis situations use Claude Sonnet 4.5; task proposals and task queries use Haiku 4.5. Experts can update the applicable rules as guidance evolves.

Deterministic stress calculations and controlled actions

Mindory works with signals such as questionnaires, heart-rate variability, sleep, resting pulse, task load and environmental information. Background stress computation uses deterministic logic. AI contributes when interpretation, personalisation or natural-language reasoning is needed.

The architecture also separates agent reasoning from application authority. Agents retrieve information and propose actions through defined tools and APIs. The application decides whether changes to tasks or other user data are persisted.

This gives AutiHD a path to extend agent capabilities while keeping control of its core application state.

Observability for individual AI interactions

CloudHero implemented OpenTelemetry logging and telemetry into Amazon CloudWatch and AWS X-Ray. User, session and request identifiers let the engineering team trace an interaction through classification, execution paths, timings and errors.

Token usage is recorded per model, component and direction. The team can reconstruct model cost for an individual interaction and investigate unexpected responses using the processing path behind them.

The rebuilt layer also incorporates Amazon Transcribe streaming for voice interactions, replacing the original batch-oriented approach.

A production AI layer delivered in 35 working days

One lead engineer and one solution architect designed and implemented the new AI application in 35 working days of engineering delivery. Improvements were continuously connected to Mindory’s production product, allowing customer testing and expert feedback to shape the architecture during development.

CloudHero used its internal agentic software-development framework for implementation, testing, code review, debugging and iteration. Senior engineers retained responsibility for architecture, technical decisions, production behaviour and customer acceptance.

Before launch, an automated 42-scenario test suite verified every hard clinical constraint covered by the project. The suite checked required system behaviour against the project’s hard constraints.

An architecture AutiHD’s developers could extend

The handover included technical documentation and a developer guide covering the request lifecycle, agents, tools, memory, stress logic, deployment, logging and extension points.

AutiHD’s team was already discussing additional task behaviour, morning and evening check-ins, more memory sources and new behavioural situations. The new architecture gave its developers defined places to add those capabilities.

“The main construct of the chat is fantastic.”

Joel Schmit, Founder and CEO, AutiHD

AutiHD and CloudHero remain in communication about potential future extensions to the product and architecture.

AWS services used in Mindory’s AI architecture

• Amazon Bedrock
• Amazon Bedrock AgentCore Runtime
• Amazon Bedrock AgentCore Memory
• Amazon Bedrock Guardrails
• Amazon Titan Text Embeddings V2
• Amazon CloudWatch and AWS X-Ray
• Amazon Transcribe streaming
• Amazon Simple Storage Service (Amazon S3)
• Amazon Relational Database Service (Amazon RDS)

AutiHD started with a live AI product and an internal development team. CloudHero supplied the production AI architecture experience needed to turn that implementation into a dedicated service its own engineers could continue building on.