Francisco MinguezData & AI

Evidence Atlas

What can I inspect to verify Francisco's work and methodology?

What you are looking at
An Evidence Atlas over the real published corpus. Each object carries Form, Provenance and Status as separate dimensions.
What you can do
Explore by topic or by form, open an Evidence Object contract, follow labelled relationships (including limits), then inspect deeper where a detail page or external demo exists.
What it means
Primary action is inspect/verify. Synthetic, research, lab, professional and academic items stay distinct. No strength scores, stars or confidence percentages.

Classified evidence only. Holdouts that were unblinded are marked holdout consumed. Explorer, Study and commercial service stay distinct.

Explore the corpus

Two proportionate views for a small corpus: by topic or by form. No enterprise filter or search theatre.

Explore the corpus
By topic

RAG & retrieval

Model selection

Release protection

Data & pipelines

Applied ML

By form

Professional Work

Research

Benchmark

Technical Lab

Demo

Academic Study

Evidence objects

Analytical pipeline modernisation

Form
Professional Work
Provenance
ProfessionalMixed
Status
Published
What it is
Professional data-engineering delivery migrating a recurring analytical process from legacy SQL Server logic to Databricks and PySpark.
Question / context
How do you modernise a brittle analytical calculation without losing traceability of business rules?
What was done
Mapped legacy logic, rebuilt on Databricks/PySpark, added quality and validation controls, reconciled outputs, and prepared handover.
What was learned
Explicit reconciliation and modular components make migration discussable; private client artefacts stay out of the public record.
What it supports here
Professional delivery experience in data-pipeline modernisation and validation discipline.
What it does not prove
Quantified ROI, client KPIs, a reproducible private environment, or production guarantees.

Where to inspect deeper

Open the professional Evidence Object

Data Quality Pipeline Lab

Form
Technical Lab
Provenance
SyntheticPublic
Status
Published
What it is
Public synthetic PySpark lab showing contracts, quarantine, reconciliation, idempotent runs, tests and CI.
Question / context
Can data-quality controls be inspected when schemas, rejected rows and reconciliations are explicit?
What was done
Built a local batch pipeline on synthetic B2B fixtures with five source contracts, quarantine, reconciliation and CI.
What was learned
Contract-first design and row accounting make accept/quarantine decisions inspectable without private systems.
What it supports here
Technical-lab method for contract-first pipelines adjacent to professional modernisation work.
What it does not prove
Employer or client system equivalence, Databricks/cloud/streaming production, or business KPIs.

Where to inspect deeper

Open the lab Evidence Object

RAG Reliability study v1

Form
Research
Provenance
SyntheticPublic
Status
Holdout consumed
What it is
Research / methodology evidence: versioned synthetic benchmark with a frozen evaluation protocol and SHA-256 evidence package.
Question / context
When should a RAG system answer versus defer under conflicting or weak retrieval evidence?
What was done
Compared lexical/dense/hybrid retrieval with a pre-registered top-1 consensus gate and a one-time unblinded holdout.
What was learned
Stricter answer/defer rules can be evaluated reproducibly on a synthetic holdout; the holdout is then consumed.
What it supports here
Methodology credibility for RAG reliability investigation. Distinct from Explorer and from the commercial F1 engagement.
What it does not prove
Production RAG performance, commercial readiness, client outcomes, or Explorer metric equivalence. Holdout is consumed; not unseen validation.

Where to inspect deeper

Open the study Evidence Object

LLM Model Selection Benchmark

Form
Benchmark
Provenance
Synthetic
Status
Holdout consumed
What it is
Synthetic benchmark comparing model/configuration candidates under a frozen evaluation protocol.
Question / context
Does any candidate clear an agreed quality bar under the same workload and constraints?
What was done
Ran a frozen synthetic evaluation; observed outcome NO ELIGIBLE RECOMMENDATION with holdout consumed.
What was learned
Eligibility can fail honestly; no recommendation is a valid decision outcome.
What it supports here
Method illustration for Compare / model-selection decision framing.
What it does not prove
Commercial validation, client results, production model ranking, or RAG-specific product proof.

Full contract on this Atlas; no separate case-study page.

Where to inspect deeper

View in selected work

GenAI Regression Gate

Form
Demo
Provenance
Synthetic
Status
Active
What it is
Fixture-first synthetic demo of a release gate blocking a critical behaviour regression.
Question / context
How can a release gate show baseline PASS → mutation → candidate FAIL → release BLOCKED?
What was done
Demonstrated a synthetic regression path with explicit fixture-first provenance.
What was learned
Gate narratives can make blocking behaviour inspectable without claiming live causal proof.
What it supports here
Method illustration for Protect / release-protection decision framing.
What it does not prove
Live GenAI causal proof, commercial validation, or a model-selection product.

Full contract on this Atlas; no separate case-study page.

Where to inspect deeper

View in selected work

RAG Reliability Explorer

Form
Demo
Provenance
Synthetic
Status
Active
What it is
Interactive companion demo with its own synthetic corpus. Subordinate to the study, not a peer offer.
Question / context
What does synthetic RAG interaction feel like when poking retrieval behaviour?
What was done
Published an interactive demo separated from the Study benchmark and holdout.
What was learned
Interactive exploration helps intuition; it must not be read as Study metrics.
What it supports here
Synthetic interaction illustration beside the Study research object.
What it does not prove
Study holdout metrics, client metrics, or a packaged commercial offer.

Where to inspect deeper

Open external demo

Predicting bank deposit subscriptions with ML and a polynomial network

Form
Academic Study
Provenance
Public
Status
Published
What it is
2024 academic study comparing classical ML baselines with a custom polynomial neural network on a public bank-marketing dataset.
Question / context
How do you evaluate a custom architecture against classical baselines with limitations stated?
What was done
Prepared data, built baselines, designed a deep polynomial network, and reported results with explicit limits.
What was learned
Baselines first, increase complexity only when justified, and keep limitations beside results.
What it supports here
Academic ML evaluation discipline under Applied ML, not a commercial delivery claim.
What it does not prove
Production banking deployment, generalisation to other institutions/periods, or client ROI.

Where to inspect deeper

Open the academic Evidence Object

Labelled relationships

Directional records. Limits are part of the graph; does not establish is intentional.

Natural next step

Inspect an object that matches your uncertainty, read its limits, then discuss a concrete problem if engagement is useful. Do not treat synthetic or study results as production proof.

Claim boundary

No fabricated clients, ROI, testimonials, validation scores or production guarantees. Synthetic material is not upgraded into client proof or a production guarantee.