Skip to main content
Skip to content
dypsis.ai
How we work

Discovery, proof of value, production. Then again.

We keep a book of rules and hold each other to it. Not a style guide, just the few a good project quietly enforces: every door has a lock, log what changed and never the value, finish it and verify it and land it green. Each round through the cycle starts from what the last one measured, and what we learn on the way goes back into the book.

Book a callSee the packages
01AIDiscoveryTwo weeksfixed fee02Proof ofvalueWeeksfixed scope03AdoptionandStagedone agreement04OperateandMonthlyrolling05TrainingpathwaysFour coursesteam priceTHEN AGAIN
CH.03 · How we work

Discovery, proof of value, production. Then again.

Every customer runs the same cycle and sees results in weeks. Each round starts from the numbers the last one produced.

Then round again, with what we measured.

DiscoveryProof of valueProductionMeasureevery customerevery round
  1. 01

    Discovery

    two weeks

    We write the problem in one sentence and look at your data. You get one prioritised backlog and a clear go or no-go.

  2. 02

    Proof of value

    weeks

    A working agent or skin on your data. Failure cases written down before anyone trusts it.

  3. 03

    Production

    installed

    On your systems, under your permissions, with a trail. We stay until it runs without us.

  4. 04

    Measure

    always

    What changed, in words and in the numbers you already track. The next round starts here.

CH.04 · Consulting

A decision with an architecture behind it.

Which layer, which runtime, which team goes first, what stays local. You leave with an architecture and a reason for each choice, then a go or no-go.

  • AI Discovery SprintTwo weeks for a fixed fee. You get one prioritised backlog and a clear go or no-go.
  • Solution architectureWe choose the model, the runtime and the connectors, and agree where your data may go and where it may not.
  • Governance designWe set who reaches what, under which limits, and what gets recorded.
AI Discovery SprintTalk to an engineer
StrategyArchitectureGovernance
decideAI Discovery Sprint
CH.05 · Implementation

We build on your systems and stay until it runs.

One team covers machine learning, generative AI, agents and skins. The first deliverable is a proof of value on your data, with the failure cases written before anyone trusts it.

  • Proof of valueA working agent or skin on your data in weeks, with the failure cases agreed before anyone trusts it.
  • Machine learningChurn, forecasting, and federated learning for when the data cannot move.
  • Generative AIRAG, agentic workflows, fine-tuning, local LLMs.
Proof of valueTalk to an engineer
Machine learningGenerative AIAI SkinAgents
buildProof of value
CH.06 · Support and maintenance

Models change. APIs move. Someone answers.

Rules drift and vendors ship breaking changes. A named engineer absorbs them before your users see them, and Atalaia keeps the estate in view for the people who decide.

  • Model and API changesWe absorb and test them before they reach your users.
  • Atalaia on the estateWe report costs, permissions and drift to the people who decide.
  • A named engineerOne person who knows your systems answers the phone.
Adoption and integrationTalk to an engineer
OperationsAtalaia
runAdoption and integration
CH.07 · Training

The skin keeps growing after we leave.

Developers learn to code with agents against your house standards. Business teams learn to work inside the assistant. Leadership learns what to govern and what to leave alone.

  • DevelopersYour developers learn to code with agents against your house standards.
  • Business teamsYour business teams learn to work inside the assistant, with the skills the skin ships.
  • LeadershipPositioning and governance workshops for the people who decide.
Training pathwaysTalk to an engineer
DevelopersBusiness teamsLeadership
teachTraining pathways
CH.04 · Where it runs

Where it runs is a decision we make with you.

The same build runs in three places. Your compliance rules, your data and the tools your teams already have decide which.

Your assistant

Inside the one your teams already use

It works inside Claude, Codex or Copilot, so nobody has to open a new application. If you do not use an assistant yet, we choose one with you.

Local models

On your hardware, when compliance requires it

Open models, installed and tuned inside your perimeter. Nothing leaves the building.

Cloud, on our terms

Through OpenRouter, only models with no data retention

Pinned to the region your data must stay in. Europe by default.

Where your data may go is settled with you before the first line of code.

CH.08 · What we need from you

We need four things from you before we start.

Every ask on this list is something only you can bring. The boxes are drawn empty on purpose.

A planning session with a client

  1. A sponsor who can say go or no-go

    One person who can decide, for two hours a week during discovery.

  2. Read access to the systems in play

    Read-only access to the ERP, the CRM and the mailbox. Nothing is written until production.

  3. One process you already suspect

    The job five people do five different ways. That is where we start.

  4. The numbers you already track

    We measure against them, not against new ones invented for the report.

CH.09 · Packages

Four stages, each with a price. Three of them end.

Start

Bring us the problem nobody has cracked yet.

We are a small team of senior specialists. We pick the right model and the right layer, and we build the least machinery that does the job. You get a call with an engineer, not a sales deck.

Book a call