Back to home

Evaluation and building

AI / GenAI

I decide with data, not fashion.

I define criteria and evals to choose models and libraries, and I stay hands-on: I build to validate ideas with clients.

Responsibilities: organising AI agents, reviewing what comes out with a critical eye, deciding with a strategic view, evaluating models and infrastructure (cloud or on-premises) and ensuring compliance (GDPR).

Pipeline stagesEvaluationBuilding

Efficiency and speed, with a critical eye

With well-organised AI agents and a minimum of critical eye, you can get very far. At the ideation stage, that is enough: what counts is efficiency and speed, with critical thinking and strategic vision.

  1. Organise

    I give the agents context, a goal and limits, so they work in the right direction.

  2. Question

    I look at what comes out with a critical eye, before accepting it.

  3. Decide

    The strategic vision is mine: what it is for, what it is worth and what to do next.

Tools change quickly. The responsibility to organise, question and decide stays.

Evaluate and build

Here too, these are the practices I use today, adjusted to each project.

Evaluate

Decide with data

Before proposing, I work out what is possible and what it costs. When there is no direct experience of the process yet, I do R&D with the team.

  • Understand what possibilities the technologies offer
  • Decide the type of delivery: cloud or on-premises
  • Gather the necessary requirements
  • Estimate effort and costs
  • Define criteria and evals to choose models and libraries
  • Take available infrastructure and compliance (GDPR) into account

Build

Show it working

I stay hands-on. Once the PoC has been awarded, I may build it myself or guide the team through it.

  • Regular deliveries, to speed up the process
  • Choose the best strategies with the team
  • Show first (diagram, mockup) when it helps the client see the idea
  • Validate each delivery with the client
  • At the end, decide with the client whether to move on to a project

Before building, I show.

The client decides better when they see the idea. So I illustrate and validate with them before any investment: a drawing, a mockup, a PoC.

  1. Process diagram

    I draw how the work is done today (BPMN) and confirm with the client that I understood it correctly.

    The client sees
    The process, step by step
    What is validated
    That I understood how the work is done today
  2. Mockup

    Wireframes and mockups of the screen, so the client can see and comment on the idea before there is any code.

    The client sees
    The solution’s screen
    What is validated
    That the idea makes sense to the people who will use it
  3. PoC

    A small working version, with the client’s data, to test the idea before investing.

    The client sees
    The idea working
    What is validated
    That it works and is worth the investment

Problems I have already solved this way

  • Scattered knowledge, slow answers

    Virtual assistants that answer from the organisation’s own information, instead of making people go and look for it.

  • Hand-made proposals, always alike

    Automatic proposal generation from tender specifications and technical product catalogues.

  • Finding similar opportunities

    Identifying similar deals through visual recognition of products.

All of them started as PoCs: a way of validating the idea with the client before investing. Several became sold projects and solutions in use.

An AI core for clinical data in European Portuguese

Context

was born with a multilingual vision, starting in Portugal, with the aim of reducing the time healthcare professionals spend on clinical records.

Constraint

Sensitive clinical data, subject to GDPR and sector regulation, and inference on local infrastructure. Added to this is the scarcity of European Portuguese data that can be used responsibly and with appropriate licences, in a market where most models and datasets are based on Brazilian Portuguese.

How pseudonymisation works

Fictional example, created to illustrate. No real data is shown.

  1. Transcription (ASR)
  2. Entity detection (NER)
  3. Pseudonymisation
Pseudonymised

Consultation on . The patient , , reports chest pain for three days. Seen at by . Continues usual medication. Follow-up booked for .

Original (fictional example)

Consultation on 14 March. The patient Joana Almeida, 67 years old, reports chest pain for three days. Seen at Vale Verde Health Unit by Dr Marta Ferreira. Continues usual medication. Follow-up booked for 28 March.

Hover over (or tap) a tag to see what was replaced.

Decision

  1. A modular pipeline of ASR, LLMs/SLMs and NER, with pseudonymisation of sensitive data before any reuse.

  2. Each model choice validated with purpose-built evals in European Portuguese, rather than generic benchmarks.

  3. An architecture ready for new languages without starting from scratch.

Result

By the users’ own estimates, a significant reduction in the time spent on note-taking. A pseudonymised database, ready for the future training of models.

My role

Led the product and the AI core for about a year, with a team of 6 people, working with stakeholders and clinical teams.

Facts and sectors

  • SeveralPoCs converted into sold projects

Skills map

Pick a skill to see where I have used it and what sits close to it.Only what belongs to this area appears here.

Applied and GenerativeAIStrategy, Adoption andCompliance

Applied and Generative AI

Strategy, Adoption and Compliance

The tools that support me

They support the work rather than define it: they change quickly and are chosen for the problem.

AI frameworks
PyTorch · Hugging Face Transformers · sentence-transformers · spaCy · TorchAudio · vLLM · TensorRT · LlamaIndex · OpenAI API · Azure AI
Development
Python · FastAPI · TypeScript · Node.js · React · React Native · SQL · PostgreSQL · Docker · Git
Tools
Claude · Claude Code · OpenCode

Shall we talk?

LinkedIn is the best place to find me.

Talk on LinkedIn (opens in a new tab)

Ask me

Answers written by me, not generated by AI.

Pick a question. The answers are mine.

Would you rather ask your agent? Connect it to me

No results.

Arrow keys to move, Enter to open, Esc to close.