Service
I design and build LLM agents that run in production, at banks, fintech companies and in healthcare. One of them routes more than 1 million help desk tickets a year at a major Italian bank, with 81% accuracy against a 75% human baseline, and cut resolution time from 5–6 days to 2–3.
An agent demo is easy to build. An agent that holds up with real traffic, real data and real compliance rules is much harder, and most stall between proof of concept and production: no evaluation, no guardrails, costs nobody can pin down.
Each agent ships with an evaluation set, a clear way to hand over to a person, and cost tracking.
Python, FastAPI, LangChain, LangGraph, LiveKit; Azure OpenAI, OpenAI, Anthropic Claude, Gemini; Qdrant, Pinecone, ChromaDB; Kafka, Docker, Kubernetes, OpenShift; Azure, AWS, GCP. If your compliance rules require it, the models run in your EU cloud tenant.
Pick one workflow with high volume, clear rules and a baseline you can measure today. An agent pays off when it beats that baseline on a metric you already track. Accuracy compared with your people is the most convincing one.
A system where an LLM decides the next step: classify, look something up, call a tool or hand over to a person. A plain chatbot only generates text. In production an agent also needs evaluation, guardrails and logging. One example is an agent that routes 1M+ bank tickets a year into 400+ categories.
You compare it with the baseline you already have. The bank ticket agent was measured against human routing: 81% accuracy vs 75%. The fashion assistant used a strict rule, where a response counted only if every item in it was correct. Without a baseline and an evaluation set, "it works" means nothing.
Yes. For banking clients the models ran on Azure OpenAI inside the client's EU cloud environment, sensitive data stayed out of the logs, and the pipelines were deployed on the client's own platforms, OpenShift included.
It depends on scope and integrations. A narrow, well-defined agent can go live fast: the IGEA voice agent was fully operational at a clinic in 2 weeks. Integrating with banking systems takes longer because of the security and compliance reviews.
Whichever fits your constraints. I have shipped on Azure OpenAI, OpenAI, Anthropic Claude and Gemini. When data rules are strict, where the model can run and under which contract often matter more than benchmark scores.
Related
Next step
Is there a workflow that eats your team's hours? We can check whether an agent can run it.
Tell me the workflow. We scope it on a free call, then a working version is live within 10 days.
Book a scoping call