Case Studies
Care
Statement
Report
Client: Mirus Australia
CONTEXT
Mirus Australia is a trusted consulting and software solution provider for aged care organisations across Australia. Their solutions include Care Statement Reports designed to monitor and communicate the life and care of residents in aged care homes. These reports give an overview of residents' overall well-being — their activities, health and therapy status, nutrition, mood, and more — and are sent to both residents and their families monthly.
PROBLEM
The Care Statement report consists of eight pages / sections, 22 sub sections and a total of 33 data points from four different systems. Usually, age care centres have between 20 to a thousand beds, and this report needs to be done for each resident every month. It takes about 1-2 hours to complete each report. Due to the work being highly manual, every month a lot of working hours is spent by senior care staff which is costly and a poor use of their time.
SOLUTION
The project aimed to automate Care Statement report generation by combining Python-driven ETL pipelines with AI-powered narrative generation, boosting efficiency and cutting operational overhead. To achieve this, we built a lean ETL flow in Python for data ingestion and transformation, then leveraged AWS serverless services for hands-off scaling.
We used AWS Lambda for all compute tasks, AWS Step Functions to orchestrate each ETL step, Amazon S3 to stage raw and intermediate data, and Amazon RDS for relational storage. This combination met the client's goal of minimizing both infrastructure footprint and ongoing maintenance.
At the heart of the automation, we integrated AWS Bedrock's LLM endpoints into our Python scripts to generate human-readable report summaries and answer ad-hoc queries—complete with built-in retries and prompt-caching to control latency and costs. We also employed advanced prompt-engineering workflows (for example, prompt-chaining) to extract structure from unstructured data and maximize the LLM's accuracy and consistency.
Despite a tight timeline and holiday season challenges, our experienced team kept the project on track. The team's proactive and pragmatic stance helped manage the project's complexities effectively. A lot of positive energy, as well as a good sense of humor, definitely helped.
RESULTS
Mirus Australia demonstrated the report to their clients and has received overwhelmingly positive feedback. We made an MVP. It works, concept proven. Roll out planning for four sites is almost complete!
Next step, LAUNCH!
AI-powered
Compliance
Client: Letzz
CONTEXT
In highly regulated industries such as iGaming, keeping up with compliance can be time consuming and costly. Resource constraints often force compliance teams to be reactive, focusing on keeping the business compliant instead of being its partners, actively involved in company strategy and sustainable growth. The available tools fail to provide significant assistance.
PROBLEM
Overwhelming changes, increasing costs with smaller or static budgets, jurisdictional knowledge gaps, the lack of efficient tools and constant firefighting - it is all straining scarce and costly resources, making teams struggle to balance compliance and growth and keeping the potential strategic value of compliance unutilised.
At the same time, there is understandable skepticism about the use of automation in compliance. The key issue with using artificial intelligence in the field of law and legislation is the level of accuracy, which must be very high.
SOLUTION
We developed a technical solution for Letzz, the first AI-powered compliance workspace. The platform continuously monitors and validates official regulatory updates, and guides compliance teams on interpretation and implementation. It helps compliance officers provide faster, more precise answers to any compliance questions the business might ask — from marketing campaigns to entering new markets.
Every query or change comes with full traceability and audit-ready evidence. At any moment, the team can see the exact source and legal basis, bringing peace of mind and confidence that the change was interpreted correctly. And Letzz tracks everything — no missed regulatory changes.
Along with a continuously and automatically updated legislative database, Letzz provides a clear overview of regulatory requirements sorted by areas and topics filtered for a specific compliance team. Instead of asking questions and keeping internal records manually, all updates — along with suggestions on how to implement them across the organization — are waiting for users inside the platform.
Letzz features industry-leading encryption, strict compliance with all major data protection regulations, user-friendly design and secure, customisable environments.
Tech Stack: Java, PostgreSQL, Spring Boot, Python, React, TypeScript, Docker, Kubernetes.
RESULTS
Already in its early stages, Letzz is saving a significant amount of time that would otherwise go into regulatory research. With Letzz, operators move from manual, fragmented, and reactive compliance toward a proactive, data-driven, and collaborative model. The result is faster market entries, reduced dependency on external consultants, and fewer audit headaches — all within a secure, easy-to-use workspace that connects compliance with the rest of the business. Letzz is currently tailored to the iGaming industry, covering ten jurisdictions worldwide (Europe, South America). With new clients joining, that number keeps growing.
Letzz raised the first investment from Henrik Tjärnström, the former CEO of Kindred Group (now part of FDJ UNITED), one of the globally leading iGaming operators.
HAIP
Client: HOOLOOVOO
CONTEXT
After the AI breakthroughs of 2022, as a tech company working for clients with complex platforms and huge datasets, we decided to focus on AI. The idea was to build our expertise and use AI to improve our own operations, as well as to be able to develop our own AI products and help our clients use the technology's advantages.
PROBLEM
Most of our engineers did not have any real experience with AI. Our non-tech people had little or no understanding of AI's concepts and capabilities. At least some of them were skeptical towards AI models, and most lacked understanding of how AI can assist them in their everyday work and life. To be able to achieve expertise and even sell it to our clients, we wanted to embrace an AI-Forward approach in our everyday business, as part of the culture across all departments within our company.
SOLUTION
Our engineers developed HAIP (HOOLOOVOO AI Playground), a platform featuring several custom AI assistants, as an easy to use internal tool for introducing and using AI across our organisation.
HAIP's AI assistants are highly adaptable - using 10+ integrated AI models, they are designed to select the ones best suited for each task and leverage their advantages. As additional AI-powered tools, the assistants offer advanced chat options, such as document upload, summarization, translation and search, etc.
Providing robust encryption, with user data not used for training of AI models, HAIP is a safe, secure and user-friendly AI resource, tailored to the unique needs of different departments and teams.
How did we do it under the hood? We integrated HAIP with our Java-based library to streamline interactions with various AI providers such as OpenAI, Anthropic, Amazon Bedrock, and simplifying AI integration across applications. Our backend relies on Java, PostgreSQL, and Spring Boot for efficient service development, including secure user authentication and authorization through a custom IAM component. For the frontend, we use React with MaterialUI to ensure consistent, responsive, and user-friendly design. Docker and Kubernetes manage backend infrastructure, providing stable, scalable, and efficient application deployment and resource utilization.
AI is rapidly evolving, and HAIP is continuously evolving too — with constant improvements that address our users' business needs in increasingly efficient ways.
Tech Stack: Java, PostgreSQL, Spring Boot, React, TypeScript, Docker, Kubernetes.
RESULTS
HAIP provided a perfect gateway for democracy in AI tools and helped our people (all our people) understand, adopt and use AI. Over 60% of our people (and growing) now habitually use AI in their daily work and feel good and safe about it.
After more than a year of practical use and constant improvements, HAIP is an ideal platform for organizations in the early stages of AI adoption.
AI archaeology
of
a legacy code
Client: DualSoft
CONTEXT
DualSOFTs an IT company with over 20 years of experience in the sports betting and iGaming industry. Their fully integrated sports betting platform — online, mobile, live, and retail terminal — is used by partners and players across multiple countries. It was built and expanded over many years. As with many legacy systems, it lacked formal technical documentation: knowledge about architecture, business rules, and decision history.
PROBLEM
Every new engineer learned the system by reviewing incomplete and potentially outdated documentation, "digging" through the code, and questioning senior colleagues. Onboarding to complex parts of the system took weeks. Business logic was buried deep within the code, scattered across various places. Answers to questions such as "what does this setting actually enable?" or "is it safe to change this?" depended on the availability of senior engineers ("bus factor" risk). Then came the time to modernize the platform. The client took action. Feeding code directly into a generic AI model does not solve the problem: in a system like this, the codebase is too large for any model to "read" all at once, so it loses context and skips parts. Even riskier is the fact that AI can hallucinate — in one of the initial attempts on a real client service, the model wrote convincing descriptions that referenced 27 functions that didn't exist in the code at all. Documentation you cannot trust is worse than no documentation at all: it leads to wrong decisions.
SOLUTION
Together with the client, we decided to test a multi-agent code analysis approach and defined the repository for our PoC. That is how the AI Code Archaeologist was born — an automated system that "digs up" knowledge from source code. A team of 11 specialized AI agents analyzes the code and generates clear, structured documentation — ranging from business rules and domain glossaries to technical details with diagrams. Eight "extractor" agents, each an expert in one area, work in parallel on isolated tasks, while the remaining three "investigator" agents assemble a story in business language from that structural data — explaining what the service does and how a request travels through the system. Additionally, the system reads git history: bug fixes reveal edge cases learned in production, while change statistics show who the real expert is for each part of the system.
The documentation is intentionally split by roles — a business section (rules and concepts, without a single line of code, for product and support teams) and an engineering section (technical details with code references) — so that everyone reads only their respective layer.
AI cannot do everything alone. The system is complex, backed by years of accumulated knowledge. Therefore, the core principle of the solution was for AI to work only where it is irreplaceable, while traditional, deterministic software handles everything that needs to be precise and repeatable. Facts are extracted by code analysis tools, not AI. Document structure is not written by AI but by a template, through which a traditional script mechanically fills in the final pages, ensuring that the same input always produces an identical output. AI only writes short business narratives about what a specific part of the system does and what its purpose is — and even there, it passes through three guardrails against hallucination: prevention (before writing, it receives a closed list of all real classes and functions and is only allowed to cite from it), verification (a dedicated validator checks every mentioned name directly against the code, and generation fails if anything is hallucinated), and write-in control (a helper script checks the output format before saving and returns a precise error description to the agent if the format is incorrect).
The system was built through spec-driven development, guided by a project "constitution" defining non-negotiable rules — and continuously tested on a real, production client service (one of the platform's core services, plus three dependent repositories), where every trial run uncovered a new type of error that was converted into a fix and an automated test.
RESULTS
As a PoC (not a finished product), the solution met all three predefined success criteria: knowledge extraction, content reliability (zero hallucinated statements), and demonstration of speed (complete documentation from a single command in under an hour — a task that would take an engineering team weeks). Tool quality is backed by 360 automated tests, 73 consistency checks, and deterministic rendering. The client received verified documentation for three user profiles (development engineers, support teams, and product/management), proof that the solution is not a one-off effort but a reusable tool, and concrete, tested insight into what AI can do with their system — and what it still cannot do — as a foundation for future strategic decisions. Five client reviewers across all three profiles rated the documentation as a good starting point. Feedback also clearly outlined the tasks for subsequent iterations. The approach can be scaled further to new repositories to cover the entire legacy codebase in a relatively short timeframe — with documentation that is easily kept up to date. We learned that trust in AI is achieved through architecture, not better prompting. Everything that can be deterministic — goes deterministic. Everything written by AI — comes with a validation mechanism. We use AI where it excels (understanding and writing); traditional software for everything else (parsing, structure, checks). Testing — on real client code. Operational knowledge not found in code — environment parameter values, logs, the impact of changes on other systems — must come from other sources. The PoC clearly mapped that boundary, serving as the starting point for the next phase.