Skip to content
Glowing Pixels
Main navigation

Where work gets stuck.

Documents arrive in one place. Decisions happen in another. Someone fills the gap by hand. We build the software, workflows and infrastructure that connect the work.

Discuss your system

From scattered documents to usable records.

Capture a document once, check it against the rules and keep it connected to the transaction. The same approach extends into billing, back-office workflows and business software.

  1. Receive

    Scans, PDFs and electronic invoices

  2. Check and match

    Extract fields, validate rules, match payments

    An uncertain match goes to a person.

    Only confident matches are booked automatically. Exceptions stay visible for review.

  3. Keep the evidence

    A booking linked to its source document

One example of how the work connects

Documents, finance and business operations

Explore the use cases

Document intake and digitisation

Scanned paper, PDFs and images become searchable, classified documents: OCR with GPU acceleration, native PDF text extraction, layout and table recognition, searchable-PDF output, embeddings for semantic search. Hardened against hostile uploads (size caps, decompression-bomb and path-traversal protection, timeouts, concurrency limits).

Invoice capture and e-invoicing

Incoming invoices are parsed (template-based and ML-based), validated against the EU standard EN 16931 and the German XRechnung and ZUGFeRD profiles, and booked. Outgoing invoices are generated as ZUGFeRD hybrid PDFs (PDF/A-3 with embedded XML) that pass schema validation.

Bookkeeping automation with evidence

Bank statements are imported and matched to invoices by deterministic rules with a documented confidence score; only confident matches are booked automatically, the rest are queued for a person. Entries follow SKR03; VAT and EÜR reports are generated; every entry links to the document that justifies it. Rule checks catch double bookings and unbalanced entries before the tax adviser does.

Back office for service providers

A self-hosted back office for a company that serves many customer sites (for example an IT provider serving medical practices): document management, per-customer boundaries, workflows and reporting, with a desktop and mobile client.

Multi-tenant business platform (ERP)

One platform serving several companies or business units with database-level tenant isolation, balanced journals, versioned service and event contracts, and on-prem or SaaS operation from the same code.

Subscription billing and payments

Subscription lifecycle as a state machine, proration, dunning policy, card payments through a payment provider, exact money arithmetic in every currency.

Membership and association management

Members, roles, fees and communication for clubs and associations, with a public site and an admin portal.

A clear path from idea to publication.

Keep the brief, language versions, review and publishing in one process. Turn existing work into the formats each channel needs, and bring the results back into the same record.

  1. Prepare

    A topic, source material and reusable brand assets

  2. Produce and review

    Text, voice and video, checked before release

  3. Publish

    Website, social channels and newsletters

The work continues after publishing.

Moderation, channel insights and independent campaign monitoring keep the record connected to what happened.

One example of how the work connects

Content and channel operations

Explore the use cases

Content operations for every channel and format

One pipeline from idea to published post across brands and languages (we run it in four: EN, DE, RU, AZ): typed content with review states, calendars and reusable creatives, published to the website and to Facebook, Instagram, Threads, Telegram and newsletter channels through the platforms' official APIs, with moderation and insights read back into the same system.

Text, voice and video production

Production pipelines that turn a topic or an article into channel-ready pieces: researched and source-checked text, voice-over from text, and short video rendered from generated scripts, each with an automated quality evaluation and improvement loop before anything goes out.

Content quality audit

Automated audit of an existing content corpus: detect weak or unsupported claims, ground them against sources, adjudicate, and rewrite in batches per language, with a re-check guard so fixes do not introduce new errors.

Independent campaign monitoring

While an agency runs the campaigns, an independent read-only monitor collects advertising and analytics data into its own record, verifies it, and reports by email. A check that could not run is reported as blocked, never as passed. Data is exportable and deleted on schedule.

Private systems. Clear operating boundaries.

Company knowledge, model access and the infrastructure underneath need explicit rules. We build within those boundaries, from private search and AI to clusters, backups and local devices.

  • Your information

    Documents, company knowledge and local inputs

  • Controlled access

    User permissions, model policies and usage records

  • An operating home

    Your premises, your VPC or hosting in Berlin

Deployment is part of the design.

Where data is processed, who can access it, and how it is retained or deleted are decisions to make before the system goes live.

One example of how the work connects

Sovereign AI and infrastructure

Explore the use cases

Private AI assistant over company knowledge

Chat and retrieval over your own documents with self-hosted models: hybrid keyword and vector search, durable memory, access control per user and tenant, usage and audit logs, tool use under policy. Works with local models or with cloud providers behind one gateway.

Model gateway with budgets

One endpoint in front of several model providers and local models: routes by required intelligence tier with no silent downgrades, spends provider quotas like budgets with virtual keys per team, blocks prompt injection and jailbreaks, caches by meaning, and traces every routing decision. Works with existing OpenAI-style clients and coding agents unchanged.

Ambient clinical documentation for medical practices

A device in the practice listens to the consultation and drafts the structured note; every sentence of the draft links to the passage it came from, the clinician confirms section by section and releases, and the note flows into the practice software through adapters for the systems' internal formats and interfaces (GDT 2.1 and system-specific). Consultation audio never leaves the practice network: no subprocessor, no vendor retention, no per-seat meter. Multilingual recognition, drafts in the practice's documentation language. No diagnoses and no billing codes by design, which keeps it outside medical-device classification. Built against §203 StGB, DSGVO Art. 9, 28 and 35, and the EU AI Act deployer duties, with the compliance research documented. Offered as a configured practice appliance, a clinic installation on the customer's IT, or managed hosting under the customer's control.

Sovereign platform operations

Kubernetes clusters run by GitOps, secrets in Vault, a self-hosted code forge with CI, self-hosted team chat (Matrix), object storage and backups. We run this ourselves across several clusters, including one in our Berlin office; we operate it in your data centre or host it in Berlin.

Local cameras, voice and building devices

Camera recording, speaker recognition and device control that stay on the premises: no vendor cloud, no subscription, open protocols where the vendor allows it, edge devices (ESP32-class) bridged into the building system.

Engineering that stays reviewable.

Agents can write code. A working engineering process also needs ownership, tool boundaries, verification and a controlled path to merge. We build the tooling around that process.

  1. Assign the work

    A defined task, one owner and clear boundaries

  2. Build with controls

    Tool policies, approval boundaries and durable sessions

  3. Verify and merge

    Evaluation, review and repository-enforced policy

A finished edit is not a verified release.

Checks and review provide the evidence for the next decision. Merge rules are enforced by the code forge.

One example of how the work connects

AI-assisted engineering with controls

Explore the use cases

AI coding agents inside a real engineering estate

Several AI agents shipping to the same repositories without stepping on each other: single-owner lanes, a four-tier escalation ladder, verification rules earned from real failures, and merge policy enforced by the forge.

Agent tooling

MCP tool servers, an autonomous development agent with typed tool policies, approval boundaries and durable sessions, code-analysis tooling, model evaluation and benchmarking harnesses, so your team can adopt agents with evidence instead of hype.

Your problem may cross these lines.

A workflow rarely ends at the edge of one tool. Bring the process, the systems around it and the point where work stalls. We can start there.

Discuss your system

Bring one workflow or operational constraint. The first 30 minutes are free, with no lock-in.

How delivery and hosting work

Every use case is available two ways: in your own infrastructure, so data never leaves your network, or hosted by us in Berlin, Germany, on infrastructure we operate ourselves.