3 min read
How we cut customer-support inquiries by 80% — and the pitfalls on the way
A real project story: an LLM support system across multiple brands, what actually moved the number, and the three mistakes that nearly sank it.
Case studyAgents
AI engineering consultancy
We build AI that ships fast and holds up in production.
An AI engineering consultancy with a decade of putting data and AI into production, from telecom to medtech. We design, build, and ship the agents, automation, and tools your business can run on, and we’re genuinely easy to work with while we do it.
ConceptProduction
Trusted by teams at
Deutsche TelekomZimmer BiometBetssonWilliam HillAbie InternationalKeepMePosted+ fintech, healthtech & telecom
What we do
Three ways we work with you.
Not a sprawling service menu. Most engagements start with a quick readiness check, turn into something we build and ship, and — if you want it — leave your own team able to carry it forward.
Start here01
AI Readiness & Use-Case Mining
Find where AI actually pays off
We start by mapping your workflows and data to surface the two or three use cases worth funding — each one costed, sequenced, and tied to a number. A plan you can act on, not a sixty-slide deck.
02
AI Product & Automation
Software that does the work
The core of what we do: agents, automation, internal tools, APIs, RAG, and full products — designed, built, and shipped to production, then made reliable enough to run your business on. (See the full list below.)
03
AI Knowledge Sharing
Level up your own team
Hands-on sessions that teach your engineers and business teams how to actually use LLMs, agents, and modern AI tooling — so the capability stays in the building after we leave.
The work itself
What we build.
AI Product & Automation is the core of what we do — and it covers a lot of ground. Whatever shape the solution takes, it ships to production and it's yours to keep.
AI agents & copilots
Systems that draft, decide, and act inside your own tools and workflows — not a chatbot bolted on the side.
Workflow automation & integration
The manual handoffs between the systems you already run, removed — so people spend their time on judgment, not busywork.
Internal tools & full products
From a single-purpose internal tool to a complete product with a frontend your team and customers actually use.
APIs & backend services
Production-grade services — typed, tested, and documented — that expose AI features for the rest of your stack to call.
RAG & knowledge retrieval
Answers grounded in your own data, with retrieval and evaluation so they stay accurate instead of inventing things.
Data pipelines, evals & guardrails
The plumbing and the reliability layer underneath it all: ingestion, measurable quality, safety rails, and monitoring.
Model- and vendor-agnostic. We pick the stack that fits the problem, not the one we’re married to. The kit we reach for most:
- Python
- FastAPI
- Pydantic
- PostgreSQL
- LangChain
- OpenAI
- Anthropic
- Llama on Bedrock
- Azure
- AWS
- Pinecone & vector DBs
- RAGAS & LLM evals
- Langfuse
- MLflow
- Databricks
- Snowflake
- Airflow
- PySpark
- scikit-learn & TensorFlow
- XGBoost
- Docker
- Pulumi (IaC)
- GitLab & Jenkins CI
- Sentry
How we work
From first call to production.
A short, time-boxed engagement: roughly six weeks to something real. Structured and built to reach production, not an open-ended discovery phase.
- 01
Scope
A first call and a short requirements pass: we pin down the one problem worth solving first and agree on a clear, measurable definition of done.
- 02
Prototype
A working proof of concept in weeks, not quarters, so your team can use it and react before we commit to full scope.
- 03
Build & harden
The production version: evaluations, guardrails, security, and the reliability work that lets it handle real traffic without surprises.
- 04
Hand over
You get a system your team owns, with the docs and training to run it — and we stay on for maintenance whenever you want it.
Who's behind this
A decade of putting AI into production.
AI Blacksmiths is small on purpose — you work directly with the engineer doing the work, not an account manager relaying messages.
It’s led by Vajk Turi, an AI engineer who has spent the last decade taking data and machine learning from research into production, for teams at Deutsche Telekom, Betsson, William Hill and Zimmer Biomet.
He’s worked nearly every role in the stack, from network optimisation and MLOps to data science and, now, leading AI automation teams. That end-to-end view, and the scars from projects that turned out harder than they looked, are what let him tell you honestly what AI will and won’t do for your business, then build the part that pays off.
The approach is simple: move fast, keep it reliable, explain it in plain language, and leave you owning the result. AI isn’t magic. It’s a tool, and the value is in wielding it well.
Vajk Turi
Founder & Lead AI Engineer
- MSc, Electrical Engineering
- ~10 years shipping data & AI to production
- Based in Malta, building for clients worldwide
- Model- and vendor-agnostic
~10 yrs
shipping data & AI systems into production
80%
fewer support tickets from one automation build
5+
industries delivered across, telecom to medtech
Days
to a working prototype, not months
Field notes
All notes →What we’re learning.
Get in touch
Let’s talk about your project.
Bring the problem, even half-formed. In 30 minutes you’ll get an honest read on whether AI is the right tool — and a first-step plan you keep either way. If we build together, it’s a fixed price agreed upfront. No meter running.