Service
Generative AI and agents that hold up with real users
Everyone wants an AI agent. A demo takes a weekend. A product that handles thousands of users without costs exploding, answers drifting or support tickets piling up is a different project.
A demo is not a product
Getting a language model to answer a question well is easy. Getting it to answer thousands of different questions, from people you didn’t anticipate, at a cost that makes sense, with answers you can defend: that’s the work. Most GenAI projects stall between those two points.
We treat GenAI like any other production system: a clear definition of success, data that is ready, evaluation before launch and monitoring after.
What we’ve shipped
- At MODO we built generative AI features for an app with 19M+ users. Contextual Search launched a year before the competition and has had zero production incidents since launch.
- At Techint we deployed a RAG assistant over 20,000+ multilingual corporate articles in 2023, before this was established practice. 80 communicators worldwide use it, and it cut search time per query by 15%.
- At SIDOM AI-assisted technical support took replies from 26.5 hours of manual resolution to 1.1 minutes on average, and handed 215 hours back to the team over the measured period.
Where agents make sense
Agents are useful when a task has several steps, clear tools and a cheap way to check the result. They are a bad fit when a mistake is expensive and hard to detect. We’ll tell you which one you have before you invest in it.
Costs and control
Token costs grow with usage, and a design that’s cheap in a pilot can get expensive at scale. We track cost per interaction from the first version, cache what can be cached, and route simple requests to smaller models. You’ll know what each answer costs before it shows up on the invoice.
What you get
A use case worth building
We check whether GenAI is the right tool for the job. Sometimes a search index or a classic model does it better and cheaper.
Answers grounded in your data
Retrieval over your documents, tickets or knowledge base, so the system answers from your sources and not from whatever the model remembers.
Evaluation before launch
Test sets built from real questions, so quality is measured and regressions show up before users see them.
Production with costs under control
Deployment, usage and cost monitoring, and guardrails for what the system should never do.
How we work
Diagnosis
The real problem isn't always the obvious one at first glance. We see the situation without the bias of being inside the organization.
Right-sized solution
We propose the simplest and most fitting solution for your specific problem. Never a bazooka to kill a mosquito.
Installed capacity
When we're done, your team has full ownership of the solution. If we keep working together, it's because you want to and we can help, not because we've got you trapped in a black box.
Related cases
Fintech
MODO
Big data processing, experimentation platform, and GenAI features for the 4+ million users of Argentina's leading digital wallet.
Global Trade
SIDOM
AI-assisted technical support for Argentina's leading global-trade platform: replies that went from 26 hours to under 2 minutes.
Industrial
Techint
RAG assistant over the institutional memory of one of Latin America's largest industrial groups, deployed to production in 2023, when almost nobody knew how to do this yet.
Further reading
How Much Does a Data Project with a Consultancy Cost? Variables, Ranges, and What Nobody Tells You
A complete guide to understanding the cost of a data project in a production environment and the variables that determine where the investment actually goes.
The difference between a PoC and an MVP (and why confusing them costs you)
Most AI projects fail because there's no clear conversation about whether you need to prove something can be done or that it actually works in production. Here's how to avoid that mistake.
Why most AI projects go nowhere
Most AI projects don't fail technically. They fail to ship. Here are four mistakes we've seen over and over, and what the projects that actually work have in common.
Frequently asked questions
Is a chatbot the right answer for us?
Often not. Many problems that look like chatbots are search, classification or workflow automation. We start from the task and the user, then choose the interface.
How do you deal with hallucinations?
By grounding answers in your own sources, measuring quality against real questions before launch, and designing the fallback for when the system doesn't know. You can't make a language model infallible, but you can make its failures visible and cheap.
Can it run without sending our data outside?
It depends on your constraints. We've built with hosted APIs and with local models; our guide on adding a local LLM to Grafana shows the second approach.
How long until something is in production?
SIDOM's support assistant took 5 months; the Techint assistant, 9 months, back in 2023. Scope, data quality and integrations drive the timeline more than the model does.