Born from a Real Problem.Built for the Hardest Cases.
MiruIQ was not designed in a lab. It emerged from a real enterprise need, where security constraints were non-negotiable and document complexity was the norm.
A Swiss Credit Bank Needed More Than OCR
Within Acosom GmbH, we were serving a credit bank client for whom we had built the credit application clerk and customer portals. Then the conversation shifted to what comes next.
The client wanted to automate the document comparison process that their teams were still handling manually. In a typical loan application, customers upload multiple documents: salary statements, identity cards, employment contracts, tax declarations, and more. The bank needed to cross-compare these documents: is the name identical across all uploads? Does the employer match? Are the calculations on a salary statement mathematically correct, or do they reveal obvious forgeries?
This was not a simple extraction task. It required semantic understanding across multiple document types, with validation logic that went far beyond basic field matching.
And then there were the security requirements. Being a Swiss financial institution operating fully on-premise, the constraints were absolute: no document could ever leave their environment. No state could be retained on the provider side; every piece of customer information was classified as highly confidential. These were not guidelines. They were hard boundaries.
We documented the requirements and added it as a backlog item. At that point, it was not even clear whether we would build, buy, or partner. What was clear: this would not be easy.
When Complex Table Parsing Met Real-World Demand
A chance encounter connected two complementary skill sets, and unlocked the technical foundation for what would become MiruIQ.
Around the same time, a connection was made with someone who had been deeply researching a very specific problem: how to reliably parse complex tables from documents using large language models. Not simple grid-based tables, but the kind found in real enterprise documents, where a footnote at the top states that trailing zeros were removed because the values did not fit on A4 paper, where columns span pages, and where formatting is inconsistent.
He had been methodically testing every available approach: fine-tuning LLMs, retrieval-augmented generation, system prompt engineering, and various hybrid strategies. For each approach, he had written comprehensive test suites across every major language model available at the time. Through this rigorous, empirical process, one approach emerged as exceptionally promising.
The critical next step was making it work under the security constraints the bank demanded. That meant local AI models, local LLM deployments on dedicated infrastructure, which required investing in high-performance hardware to run larger models entirely on-premise, without any cloud exposure whatsoever.
With one side bringing deep expertise in streaming architectures and distributed systems, and a client that genuinely needed this, the decision was straightforward: build a proof of concept that adhered to every security requirement from day one.
The Proof of Concept Was a Success
What started as a focused experiment became the foundation for an entirely new platform.
The POC delivered exactly what the client needed, and they approved it. From that point forward, the mission was clear: take this validated approach and build it into a production-grade platform that any enterprise could use.
What followed was an intensive period of development. Over the course of approximately 18 months, the architecture was rebuilt three times. Each iteration brought the pipeline design closer to what was needed: Apache Flink-based streaming jobs that could isolate state not just per customer, but per individual pipeline execution. This level of isolation was not a nice-to-have; it was a direct consequence of the security model.
The result is MiruIQ as it exists today: a platform we built to be the most flexible and most secure document automation solution we know of. Not because security was bolted on afterward, but because it was the founding constraint that shaped every design choice from the very first line of code.
Built Different, Because It Had to Be
Every capability in MiruIQ traces back to a real requirement, not a feature roadmap.
Complex Table Extraction
Handles the edge cases that break other tools: truncated values, multi-page tables, footnotes that redefine column semantics, and inconsistent formatting across document types.
Semantic Extraction
Goes beyond layout-based OCR to understand what a document actually means. Extracts structured data from long, complex documents by interpreting context, not just coordinates.
Cross-Document Validation
Automatically compares fields across multiple uploaded documents, matching names, employers, dates, and calculations to detect inconsistencies and potential fraud.
Security by Design
Fully on-premise deployments, no state retention on provider infrastructure, and pipeline-level isolation. Security is the architecture, not a feature toggle.
Swiss Infrastructure
Hosted on local infrastructure with no cloud dependency. Local AI models run on dedicated hardware; your documents never leave your environment.
Pipeline Architecture
Apache Flink-based streaming pipelines with per-execution state isolation. Scalable, auditable, and designed for regulated industries from day one.
Ready to Experience MiruIQ?
Whether you need complex table extraction, cross-document validation, or a fully on-premise deployment, we would love to show you what MiruIQ can do.