Lucian MurguResearch · Engineering · Writing

Engineering

Engineering.

I design and build AI systems, evaluation pipelines and agentic workflows, and validate them with checks that can say no.

It grew from a decade of production work — institutional backends and numerical engineering software — where data, permissions and calculations had to be right before release.

Capabilities

  • AI & Agentic System ArchitectureBounded agent missions and multi-step workflows.
  • Evaluation & Reliability EngineeringValidators, seeded failures and failure analysis.
  • Model Interpretation & VerificationBenchmarks with stated limits; tested explanations.
  • Backend & Data EngineeringAPIs, data models, geospatial data, workflow state.
  • Scientific & Numerical ComputingNumerical methods, simulation, preregistration.
  • Technical Governance & ControlsAcceptance gates, vetoes, audit trails.

Technology

  • Python
  • PyTorch
  • FastAPI
  • NumPy
  • scikit-learn
  • PostgreSQL / PostGIS
  • Next.js
  • Java
  • Spring
  • Oracle SQL
  • C/C++
  • Fortran
  • Docker
  • Playwright

Used in the projects on this page and in professional roles.

Selected systems & applications

  1. Platform · AgriculturePre-pilot · 2026

    AgroUnu: Platform for Crop Farms

    An evidence-to-decision platform for Romanian crop farmers: mapped fields, crop cycles, and purchase-to-dispatch traceability for every event.

    My role Design and build: data architecture, API, web app, review process.

    • FastAPI
    • PostgreSQL + PostGIS
    • Next.js
    • OIDC · tenant isolation

    Case study

  2. Application · Medical imagingResearch prototype · 2026

    Clarity — Diabetes Evidence Workbench

    A research workbench for diabetes evidence and expert review; its retinal classifier reached ROC AUC 0.942 on a 400-image benchmark. No clinical validation.

    My role Design and build: API, review workflow, training receipts, evaluation.

    • FastAPI
    • PyTorch · DINOv2 head
    • scikit-learn
    • Expert review

    Private repository · case study not yet published

  3. Scientific engineering · Structural analysisOct 2026

    Preregistered Simulation Harness

    The engineering behind the Jacobian-pullback study: a checked solver, analytic derivatives, a hash-locked protocol and a clean-room reproduction.

    My role Sole author: simulator, protocol lock, runner, reproduction check.

    • Python · NumPy
    • Analytic Jacobians
    • Hash-locked protocol
    • Reproduction check

    Read the study

  4. System · Evaluation & reliabilityPre-pilot process · 2026

    Agent Reliability & Evaluation

    Coding-agent missions checked by validators, seeded contradictions and a separate evaluator agent with veto power; its first mission was rejected twice.

    My role Mission contracts, evaluation policy, builder/evaluator separation.

    • Evaluation
    • Seeded contradictions
    • Critical vetoes
    • Git & state checks

    Case study

  5. Production software2015–2026

    Production Systems for Institutions & Engineering

    Contributions to team-built systems: the European Commission’s SEP grant-evaluation system, the EEAS Security Portal, and Bentley’s bridge-design software.

    My role Data models, permissions, workflow state, releases; numerical fixes.

    • Java · Spring
    • Oracle SQL
    • C/C++ · Fortran
    • Workflow & permissions

    Experience

More engineeringAll systems, platforms and professional work

Diagrams are drawn from documented architecture; the phantom is the project’s own synthetic training image. All repositories are private.