AI consulting · MANDGIE AB · Stockholm
AI consulting that ends with your team owning the system.
Most AI initiatives die in the gap between a promising demo and a system people rely on every day. I work in that gap: I build AI agents and LLM automation inside your company, together with your engineers and the people who will actually use them — and I leave behind a team that can run, evaluate and extend what we built.
What I build
Agents that do real work in your real systems.
-
AI agents wired into your stack
Tool-calling agents connected to the systems you already run — CRM, data warehouse, email, internal APIs — that take over manual gathering, drafting and coordination work. Built with the OpenAI and Anthropic APIs, MCP, and RAG where it earns its place.
-
LLM automation in existing workflows
Not a new tool your team has to adopt — automation embedded where the work already happens, so using it takes less effort than not using it.
-
Evals, guardrails & monitoring
The part that separates production from a demo: an eval suite grounded in your real cases, guardrails around what the system may do, and monitoring that catches drift before your users do.
-
Advisory & team training
For teams building themselves: reviews of agent architectures, and hands-on training in evals, integration patterns and failure modes — the things I've learned shipping LLM systems since 2018.
Recent results
Measured in hours, not vibes.
An affiliate network — 4–5 hours a day back, per key account manager
Key account managers spent a large share of every day on manual data and information gathering across internal systems before they could do their actual job: advising partners. An agent now does that gathering for them. Each account manager saves 4–5 hours a day. Read the full case study.
A dental care company — one agent the whole business can ask
A company-wide business-insight agent, grounded in the company's own data, that lets anyone — not only analysts — ask what is going on in the business and get a straight answer.
Clients are anonymized here; I'm happy to go deeper on either engagement in a conversation.
How an engagement works
Weeks to production, then handover.
-
Pick one workflow that hurts
Not "an AI strategy" — one concrete, measurable workflow where hours are visibly being lost. We define what success looks like in numbers before anything is built.
-
Build on site, with your team
I implement alongside your engineers and the people who will use the system. The integration patterns, the eval suite and the failure modes get learned by your team while the system is being built — not handed over in a document afterwards.
-
Hand over ownership
The evals, the monitoring and the playbook stay with you. The engagement is designed so you don't need me afterwards — which is exactly why clients come back for the next workflow.
Why me
Eight years of production, three vantage points.
I've built and run AI systems as an operator inside a large company, as a platform lead in a scale-up, and as a founder with my own money on the line:
- Spotify (2018–2022): led the four-person team behind the internal intent model — transformer NLP adopted from BERT within months of its release, still in production today.
- SeenThis (2022–2024): built the data platform end to end — 10,000 events per second in real time — plus the ML the business priced its product on.
- Saltfish (2024–2026): co-founder & CTO of an AI platform for B2B SaaS; raised a $730K pre-seed led by Antler; exited via share sale in 2026.
The full history is on the home page, and I write about what production teaches in eight years of production NLP.
Start here
Describe the workflow. I'll tell you honestly if AI fits.
One email is enough — what the work is, who does it, and roughly how much time it eats. I'm based in Stockholm and work on site. I read and answer every email myself.
magnus@mandgie.comOr reach me on X: @mangefriberg