Skip to main content
IRApplied AI
EngineeringContact

Isaac Reis · Applied AI Engineering

Applied AI engineering for complex industrial workflows.

I build bounded, testable systems that connect AI models to deterministic engineering tools, traceable evidence and expert approval.

For industrial R&D teams, pilot owners, research organisations and European consortia.

Explore the demonstrators Discuss a pilot or consortium
IR / VALIDATION LOOP / 0138.7223° N · 9.1393° W
  1. 01Model proposalGenerated
  2. 02Contract checkPassed
  3. 03Simulation / testMeasured
  4. 04Expert decisionRequired
TraceableReproducibleReviewable
Bounded AIDeterministic checksTraceable evidenceHuman approval

The simplest method that survives the evidence gate wins.

Demonstrators built around real evaluation constraints.

These projects explore CAD and simulation, energy interoperability, manufacturing inspection and evaluation-led clinical research. Each claim is limited to its current evidence state.

01

Industrial CAD · Simulation

Application context: AI-BOOST Challenge 2

CAD-Agent

From design intent to traceable geometry and simulation evidence.

A working vertical slice that turns typed intent and controlled drawing profiles into constraint-checked routes, STEP geometry, tetrahedral meshes, first-order FEA outputs and review artefacts.

Evidence today
The described software path has passed current repository checks. Its benchmark, geometry and numerical results remain synthetic and await independent CAD, piping and FEA review. Self-assessed maturity is TRL 3; official effort and package-size KPIs are unmeasured.
Next gate
Seeking piping, rotating-equipment and CAE reviewers, representative 2D-to-3D workflows and measured human baselines.
  • Repository checks passed
  • Synthetic evidence
  • Domain review pending
  • Official KPIs unmeasured
02

Energy · Interoperability

Application context: HEDGE-IoT Open Call 2

FlexAgent

Reproducible reserve-exchange evidence at a simulator boundary.

A deterministic test workflow for mFRR and aFRR exchanges, with a technology provider emulating the balancing-service-provider boundary. It models timing, setpoints, metering, trace correlation and communication faults.

Evidence today
Current runs use synthetic scenarios and standard-neutral traces. No IEC 62325 codec, validated pilot profile, simulator adapter or R&D Nester connection is claimed.
Next gate
Seeking simulator-contract access and review from IEC 62325, flexibility-market and energy-interoperability specialists.
  • Repository checks passed
  • Synthetic evidence
  • Domain review pending
03

Manufacturing · Edge vision

Application context: FutureProofTextiles Open Call

Hybrid Edge Inspection

Local defect screening with visible evidence and operator control.

A one-station POC combining conventional anomaly/localisation candidates, a local vision-language reviewer and append-only operator corrections. Final quality disposition remains with the operator.

Evidence today
The workflow and audit controls are implemented. Public and synthetic diagnostics also showed that the current VLM is not reliable as a sole detector. No factory accuracy, throughput or ROI claim has been established.
Next gate
Seeking a textile manufacturer or QC partner to define capture, defect taxonomy and a roll/batch-isolated pilot evaluation.
  • Repository checks passed
  • Public-proxy evidence
  • Domain review pending
04Partner-led research

Clinical datasets · Research only

Evaluation-led clinical dataset completion.

Public-proxy work on GDC and GENIE found selected conventional tabular methods stronger than the tested cellwise medical-LLM role, so the LLM path was not promoted. A synthetic-mask CUDA imaging integration has run; no clinical imaging evaluation has.

Not for clinical use. Current results do not establish clinical efficacy, safety, privacy, fairness, regulatory compliance or challenge KPI attainment.

How to read the evidence labels
Repository checks passed
The described path is implemented and passed the current repository checks; detailed evidence is available for review.
Synthetic evidence
Results use constructed inputs and are not deployment estimates.
Public-proxy evidence
Results use a different public dataset and do not establish organiser-data performance.
Domain review pending
Qualified independent review has not yet occurred.
Official KPIs unmeasured
The programme's required evaluation protocol has not been completed.

Evidence before adjectives.

I use the simplest method that survives a pre-declared evaluation. AI earns a role only when it adds measurable value to the workflow.

  1. 01

    Frame the decision

    Identify the user, workflow, inputs, constraints, failure modes and authority boundary.

  2. 02

    Freeze the contract

    Define typed interfaces, assumptions, baselines, success criteria and the evidence required for every claim.

  3. 03

    Build the vertical slice

    Connect probabilistic models to deterministic tools, APIs and operator workflows.

  4. 04

    Test honestly

    Use held-out cases, serious non-AI baselines, provenance, fail-closed gates and explicit uncertainty.

  5. 05

    Validate with experts

    Measure quality, effort and rework in a relevant environment. Promote claims only after qualified review.

Working principle

A failed hypothesis is useful evidence. It prevents expensive technology from being promoted into the wrong role.

From technical uncertainty to a defensible pilot.

Product engineering, evaluation and deployment are treated as one connected technical problem—not separate hand-offs.

01

Applied AI product engineering

Structured extraction, tool-using workflows, geometry and simulation integration, computer vision, tabular ML, APIs and operator interfaces.

02

Evidence and evaluation engineering

Typed contracts, provenance, strong baselines, held-out protocols, reproducible reports, correction lineage and human approval gates.

03

Interoperability and constrained deployment

Version-bound interfaces, trace correlation, failure testing, edge and offline execution, and integration with existing technical workflows.

Ways to work together

Focused engagements with explicit exit evidence.

01

Feasibility & Evidence Sprint

Freeze the workflow, KPI contract, baselines and evidence boundary, then build the smallest useful technical slice.

02

Demonstrator-to-Pilot

Integrate a working demonstrator into a relevant environment and measure quality, effort, rework and operator impact.

03

Interoperability & Assurance

Build typed interfaces, traceable exchanges, deterministic checks, failure evidence and explicit approval boundaries.

Founder-led, based in Portugal

Senior AI engineering with direct founder involvement.

I'm Isaac Reis, an AI engineer with seven years of experience building production-oriented AI and tool-using systems.

I founded this company to turn technically ambitious ideas into bounded, inspectable workflows that industrial partners can test. Every engagement is led by me.

Where qualified judgment is required—piping and CAE, energy-market rules, textile quality or clinical validation—I work with domain partners and keep their approval authority explicit.

Best fit

Industrial R&D teams, pilot owners, technology providers, research organisations and consortium coordinators looking for a technically self-sufficient SME contributor.

05 / Start a conversationPortugal · Working across Europe

Bring a real workflow and its constraints.

If you have a technical process, validation environment or consortium gap, let's determine where AI genuinely belongs—and what evidence would make the result credible.

Discuss a pilot View GitHub

Particularly interested in industrial CAD and simulation, manufacturing inspection, energy/IoT interoperability and carefully partner-led regulated research.