---
title: "Aura2 | Dylan Tullberg"
description: "Defining and building Aura’s AI experience, long-term memory, evaluation, privacy, and dependable delivery at Aquablu."
url: "https://tull.ink/work/aura"
---

[← All projects](https://tull.ink/work)

Aquablu · 2026–present

# Aura2

Turning business context into useful, dependable assistance.

Operational questions cross customers, connected systems, servicing, and knowledge. The challenge: bring that context into one conversation without losing evidence, permissions, or human control.

My responsibility

I defined Aura’s AI experience, built its capabilities with agents, and developed the workflows behind dependable team delivery. The wider platform scope was collaborative.

## One question, connected context.

Explore four illustrative tasks, from comparing usage to preparing a recurring brief. Each answer keeps its source and scope visible.

![Aquablu REFILL+ Series 2 in an AI-composited workplace scene](https://tull.ink/work/aura-device-scene.webp)

Health Usage Servicing

Across the business

1.  Systems
2.  Customers
3.  Service
4.  Commercial
5.  Knowledge

**Aura AI Assistant**Illustrative example

Explore another examplePrioritize servicingCompare usageFind guidanceRemember and repeatCurated examples, not a live chat.

You

Which locations need servicing, and what is already scheduled?

Aura

Two locations need attention. Demo HQ is overdue and has no visit scheduled.

Fictional servicing results

Location

Status

Planning

Demo HQ

Overdue

Not scheduled

Demo Studio

Due soon

Scheduled

Example source: servicing records · Scope: two demo locations

One platform. Scoped perspectives.

-   **Aquablu staff**Fleet-wide operations
-   **Partners**Their customers and systems
-   **Clients**Their locations and systems

Illustrative workflow with example data.

## The decisions behind the experience.

### Start with the work, not the data source.

I defined and built the AI experience around user intent and organizational scope, connecting operational records, analytics, and knowledge.

### Implementation detail

Support demand informed capability priorities. A useful answer can depend on a customer, location, device, service record, and guidance held elsewhere. My work extended into Go APIs, React interfaces, data-quality detection, and planning.

### Let Aura remember, without giving up control.

Starting every conversation from scratch makes an assistant feel like another tool to manage. I built Aura’s memory so it can carry useful context forward: how someone prefers to work, what happened before, and how customers, locations, systems, and service history relate.

### Implementation detail

Modeled after Graphiti, Aura uses the same temporal, entity-linked memory approach with semantic and keyword retrieval. Memories keep their source and history, and people can see, correct, forget, or restore them. Because memory can contain personal information, I treated it as governed PII with identity and organizational scope, GDPR-oriented retention and erasure controls, and content-free operational logging. The wider assistant experience includes explicit AI identification, human review of proposed actions, and a documented EU AI Act assessment, with external legal review retained as a launch condition.

### Make trust testable.

I identified the need for evaluation and built the system to test answers, permission boundaries, and tool use before release.

### Implementation detail

Adversarial cases, calibrated model judges, statistical gates, and release evidence distinguish synthetic tests from checks against the deployed assistant. Privacy, audit, cost controls, and data-quality detection support the experience.

### See the observability interface

Cost per successful run, measurement coverage, failures, and latency make the experience inspectable.

![Aura observability interface with cost per successful run, measurement coverage, failure rate, latency, and charts of daily cost and token behavior.](https://tull.ink/work/aura-observability.png)

Development interface, August 2026. Figures show test data.

### Make recurring work reusable, not automatic by default.

I built reusable skills and routine proposals with explicit scope and confirmation boundaries.

### Implementation detail

Reusable instructions reduce the need to reconstruct the same workflow. People review what a routine will run and when before confirming. Missing information, restricted access, and proposed actions remain distinct states.

### Build the delivery safeguards into the workflow.

I developed coordinated, agent-driven development workflows so automation and quality checks travel with the work.

### Implementation detail

I optimized test setup and CI execution, built a shared change-aware validation runner, and extended shipping automation with pipeline monitoring and deployment-drift checks. This extended an existing, collaboratively developed engineering workflow.

## An experience and a way to deliver it.

The work established a connected assistant, long-term memory, evaluation, privacy and operational controls, and repeatable development checks. Product decisions and implementation stayed part of the same responsibility.

**Continuing direction.** Broader operations, commercial activity, and financials are part of Aura’s expanding vision, not a claim that every capability is available today.

Go · React · TypeScript · PostgreSQL / AlloyDB · Vertex AI · MCP · OpenAPI

[Explore more work](https://tull.ink/work)[Discuss a role or engagement](https://tull.ink/#contact)