Projects

Private AI in practice.

Two current public projects and three anonymous business cases. We describe visible, documented work while keeping confidential client details out.

Current projects

Working interfaces connected to real sources.

Current project

Challenge: Reliable MPN information spans many sources. People need a clear way to ask questions without losing sight of where an answer comes from.
Solution: Magic retrieves relevant material from its knowledge base and displays the sources used in its answers. An administration area covers source ingestion and operational information.
Why local?: Its knowledge base and model infrastructure are designed for a managed environment without requiring a public model API.
Result: The project interface supports questions and source review. Development is ongoing.
Technology: Document indexing, retrieval, source references, model service and administration interface.

Current project

Challenge: Public records, information requests and council documents are spread across sources. Finding the relevant material takes time.
Solution: The site brings documents together through search, topics and an assistant that links back to its sources. Visitors can inspect the underlying material themselves.
Why local?: Document processing and search can be managed on dedicated infrastructure; source links keep the information verifiable.
Result: Visitors can locate public documents and ask questions about the available material. Interpretation remains with the reader.
Technology: Document access, retrieval, source links, research interface and AI assistant.

Anonymous client cases

Anonymous client case · Delivered solution

Challenge: Email and document exports in several formats needed to become usable in Dutch without repeating the same manual steps.
Solution: A batch pipeline detects and filters input, extracts content and translates it through a managed model endpoint.
Why local?: Documents can stay within a managed processing environment instead of going to a public translation service.
Result: Translated output is organised for review. Progress and processing can be tracked in a dashboard.
Technology: Batch processing, format extraction, an OpenAI-compatible local model endpoint and progress monitoring.

Anonymous client case · Delivered solution

Challenge: Sensitive recordings containing Dutch speech and regional dialects needed to become reviewable text.
Solution: An offline batch pipeline processes audio, distinguishes speech from music and tracks progress. Interrupted jobs can resume.
Why local?: Audio does not have to be sent to a public transcription service for this workflow.
Result: Outputs include timestamped segments for human review, with processing and provenance logged.
Technology: Whisper large-v2, dialect recognition, segment timestamps, repetition filtering and an operational dashboard.

Anonymous client case

Challenge: Financial records exported in several formats require time-consuming fact finding, with every observation traceable to its source.
Solution: A local Quick Scan reads CSV, Excel and SQLite files, discovers each structure, runs scoped queries and checks findings against the source data.
Why local?: The analysis can run fully offline, keeping confidential case material within the organisation's environment.
Result: The workflow prepares a factual report with points for review and source citations. Decisions about fault, liability or next steps remain with the responsible professional.
Technology: Local language models, scoped data tools, verifiable query log, source citations and human review.

Contact

Tell us how the process works today, what data is involved and what a useful outcome would look like.