A managed IT services provider running Zammad as their helpdesk — multiple clients, each with their own ticket history and recurring issues.
Every new ticket meant a technician manually searching past cases, browsing vendor documentation, and composing a first-response from memory. Slow, repetitive, inconsistent.
A draft answer in Zammad as an internal note — in Swedish — within seconds of ticket creation. Zero manual lookup.
45-word Google query → Apify scraper → Claude Sonnet strips boilerplate → pgvector.
Queries per-client pgvector table of resolved tickets. Llama 3.2 synthesises both sources into a Swedish draft.
Cloud APIs used only where local models fall short — keeping support data on-premise, costs predictable.
Zammad, N8N, Ollama, and pgvector all run on-premise. No customer data leaves the server unless a specific inference task requires it.
Draft posted before a technician opens the ticket — drawn from historical data and live research simultaneously.
The agent searches that customer's own resolved tickets first. Known solutions surface immediately.
When no past ticket matches, it scrapes current docs, summarises, and stores for next time.
Translated locally via Ollama. No cloud translation cost, consistent professional tone.
Custom AI automation for IT teams and managed service providers across Sweden and Northern Europe.
Sweden's longest unsolved murder — the 1986 assassination of Prime Minister Olof Palme. Decades of witness statements, police reports, parliamentary records, and investigative journalism exist across fragmented archives.
No single tool lets a researcher ask "Who was at Tunnelgatan that night?" and receive a sourced, cross-referenced answer drawn from the full corpus — not a keyword match, a semantic one.
Build a forensic RAG platform that ingests, embeds, and cross-links the entire available document base — then exposes it through a chat interface built for investigative reasoning.
N8N pipelines pull from Internet Archive and wpu.nu, chunk documents, and pass them through OpenAI embeddings into pgvector — with entity/relationship extraction on every document.
FastAPI serves a forensic chat interface, a connections graph, and a timeline — all drawing from the same pgvector store with full source attribution.
Ask questions in plain language. Every answer is cited to source documents with page references, confidence levels, and conflicting accounts flagged.
Entity-relationship map of persons, places, organisations, and events extracted from the corpus. Filter by date range, source type, or confidence.
Chronological view of documented events sourced across all ingested material — cross-referenced against known official accounts to surface discrepancies.
A multi-agent N8N architecture for anomaly detection — surfaces patterns, contradictions, and statistical outliers across the document base that manual review would miss.
Crime scene fieldwork at Tunnelgatan, Stockholm — original observations integrated as primary source annotations within the platform.
Active development. The Engström designation — the last official suspect — was formally retracted in December 2025, making independent investigative tooling more relevant than ever.