MindWareHouse · Founded by Víctor Hernández

Human-directed AI systems that show their evidence.

We turn documents into data, clean up and consolidate the data you already have, and research with cited sources, using AI workflows where every result carries its evidence and a person reviews it.

For NGOs, newsrooms, research and policy teams, and small and mid-size companies.

For now, requests are received by LinkedIn message.

01 / MindWareHouse

MindWareHouse: the platform underneath everything else.

MindWareHouse is not just the company name — it is the platform, in active development, underneath everything else on this site: a governed AI "organization" of specialized agents working under one person's direction. What is shown below is what runs today: governed memory with provenance (record who said what, and preserve disagreement instead of averaging it away), a deterministic code auditor, and QA that verifies what a person can actually see and use, not just what the code claims. Components that do not yet run end to end are not shown.

Stage, stated plainly: A code-grounded scan of our internal planning documents found 60 named components: 16 have code that runs, or ran and was observed running; 44 are not there yet (39 are proposals, diagrams, and specs; 5 have code that is not yet confirmed to run). (as of Sep 28, 2026) Each component below shows its maturity on its own card.

Inside MindWareHouse

3 platform components: governed knowledge, code verification, and runtime QA. Each tile opens its own brief.

02 / Why we exist

Power should be legible, not theatrical.

Most AI that reaches organizations decides without being accountable: it doesn't say where something came from, how sure it is, or who answers for it. We build the other way around: evidence first, measured uncertainty next, and a person who decides at the end.

03 / Built on MindWareHouse

Systems, not demos.

Products built on the MindWareHouse platform, each with its maturity stated: what works and where its limits are. Platform components are in section 01.

04 / A live interface

A real interface. Visible decisions.

A live public surface: the VeraVadis showcase.

Functional platform · live public showcase

VeraVadis → HelpVenezuela

An intelligence engine that shows its evidence. Its showcase publishes on an ongoing basis.

VeraVadis separates what is claimed from what is known: provenance of every source, independent sources (a source counts as independent only if it does not originate from the same statement), measured uncertainty, and what remains unconfirmed. HelpVenezuela is its operational showcase, a bilingual publication on Venezuela organized around scenarios rather than headlines: a set of open questions, each with its options, dated evidence, indicators, upcoming milestones and revision history, plus a column of the events that move them. Estimates are given in words, not percentages; an AI analyst produces them, an AI editor checks them, and the page says they have not been reviewed by a person and are not calibrated forecasts. Publishing has had documented interruptions; it is not a guarantee of uninterrupted daily activity.

  • 4,607 briefs published, each in English and Spanish (as of Nov 11, 2026)
  • 453,353 news items collected · 203,965 entities (as of Nov 11, 2026)
  • Images always labelled "AI-generated editorial image"

Every figure carries its date; news counts move as duplicates are removed.

Open helpvenezuela.org ↗
HelpVenezuela live homepage HelpVenezuela on a phone

05 / Method

01

Preserve evidence

Provenance, chain of custody, and the difference between six outlets repeating one source and six independent confirmations.

02

Measure uncertainty

Calibrated scoring, dissent, contradiction, and honest "unknown" states.

03

A person decides

Approval gates, audit trails, and reversible action.

What went wrong, and how we fixed it

We publish our mistakes because a system that never fails in public is a system nobody is measuring.

  • A "verified" seal that nothing had verified. An early prototype page of one of our products displayed an "MWH Verified" seal with no verification behind it. We retracted it in August 2026, and every amount that product shows is now labeled as an estimate.
  • An automated judge that was mostly wrong. We measured a fast model for deciding whether two claims contradict: only 5 of its 29 alerts were right (about 17%; small pilot, 40 randomly sampled claim pairs) (as of Sep 20, 2026). We stopped using it for that. A separate pilot found it does much better on a narrower, checkable task — matching the same entity across mentions: 13 of 15 correct, about 87% (small pilot, n=15) (as of Sep 20, 2026). That is not wired into production either; both remain measurement-only pilots.
  • "Five voices" that were five outlets covering the same topic. Topic is not corroboration. Now only an independent source backing the same fact counts.

06 / Operating model

Architecture is also resource allocation.

All of this runs on open-source software, a 2017-era workstation, and a modest monthly budget for model subscriptions and APIs. Not austerity as spectacle: operating discipline. Human attention is reserved for judgment.

SoftwareOpen-source stack
Computei7-7700 · 32 GB RAM
Model accessSubscriptions and metered APIs
PracticeParallel agents under human direction

07 / Leadership

Víctor Hernández builds at the boundary between strategy, systems, and service.

Technology entrepreneurship, university teaching, operations research, humanitarian leadership, civic organization, and Catholic mission work. He has directed companies, volunteer teams, and community networks under severe constraints.

The common thread is orchestration: understand the mission, assemble the right people and tools, make responsibility explicit, and turn scarce resources into a working system.

08 / Work with us

Three concrete services.

Documents to data

From documents to usable data

PDFs, scans, forms and reports turned into clean spreadsheets or databases. Every value can be traced back to the page it came from, and anything that could not be read is flagged, not guessed.

See the service on LinkedIn ↗
Data cleanup

Data cleanup and consolidation

Scattered spreadsheets and records merged into one consistent dataset: duplicates resolved, formats standardized, and every rule applied written down so the result can be checked.

See the service on LinkedIn ↗
Research

Research and monitoring with cited sources

One-off research or ongoing monitoring of a topic, a market or a region, in English and Spanish. Every statement links to its source, and what is still unconfirmed is said to be so.

See the service on LinkedIn ↗

For now, contact is through LinkedIn.