Iván de Jesús Díaz Navarro
I build and run the systems, not just the reports.
Data that doesn't lie.
Data Analyst · Business Intelligence · Operations automation
I design, build and run the systems a business actually works on: multichannel inventory, quoting tools, customer-facing agents, executive dashboards. I do not hand over a report and leave — the system stays running, with backups, rollback and a written procedure.
New video introduction · coming soon
Systems that keep running
Backup, rollback and a three-step verification
Local AI when the data is sensitive
Conversations and bank statements that never leave
None of these figures is an estimate: the first three come from counting the cases published right here, and you can verify them by opening them.
26
published case studies, each with its context, its decisions and its results
11
of them are systems running in production for real businesses
6
public sites and live demos you can open from these cases
0
open ports: ~60 containers on my own server, behind an encrypted tunnel
The fourth number is explained in full in How I work.
This is not a slogan: these are rules written into the code of these systems.
Coverage with no sales is reported as empty, never as zero.
A zero would read as «critical» and send someone out to buy what nobody is asking for.
With fewer than three observations, the dashboard says «not enough data».
A percentage computed on two data points is a number shaped like a truth.
A 2xx proves nothing.
Every write to an external channel is read back and compared; if it does not match, it is marked as failed.
Each one opens with the real problem that caused it. All six are published with their decisions, their figures and what I learned building them.
Three isolated inventories, and none of the three was the right one.
A central append-only ledger decides what stock each channel sees. 796 listings audited, 1,742 movements recorded and a 90-day query answered in 3.5 ms.
Read the full case arrow_forward512 units left the warehouse. Inventory deducted 408.
A PWA that stores events instead of balances. Replaying 3,536 real events proved the fix worked: 512 = 512, and not a single box in negative.
Read the full case arrow_forwardLocal AI or nothing: no conversation and no catalogue leaves the machine.
One engine serves customers of two different businesses over WhatsApp, with language models running on my own server. Zero conversations sent to external AI services.
Read the full case arrow_forwardA public link that computes prices is, by definition, a link that can leak how they are computed.
The customer builds their own package on a private link and the price is computed only on the server. A runtime guard blocks any cost, margin or supplier from leaving.
Read the full case arrow_forwardIn multi-tenant systems the question is not whether code will leak data between tenants, but when.
Isolation is enforced by PostgreSQL Row Level Security, not by the application. Money in cents, a single role allowed to bypass the rule, and a test that checks the voucher never leaks supplier data.
Read the full case arrow_forwardWhat you learn from a lost bid is worth as much as what you learn from a won one, and it is the first thing to disappear.
Thirteen phases built, 67 automated tests passing and 23 numbered design documents. It is the one case with a public live demo anyone can open.
Read the full case arrow_forwardEach one comes with the published case that backs it. No case, no claim.
From raw data to a dashboard that reads without a translator — including the uncomfortable part: stating clearly which questions that data cannot answer.
Inventory, messaging, invoicing and reporting that stop being done by hand — with a written limit on how far it is right to automate.
Web apps and installable PWAs the business owner runs on their own, handed over with a tutorial and without dependencies that expire.
A measured diagnosis of a live site or system: what is actually broken, what it costs the people using it, and what gets fixed first.
If you are operating on data you do not quite believe, or on a system nobody knows how to deploy, that first conversation is free.
The tools that show up in the published cases. Not one more.
Cleaning and analysis with Pandas, re-runnable pipelines, optical reading and browser automation.
Star schemas, append-only ledgers, Row Level Security isolation and queries that answer in milliseconds.
Executive dashboards where the narrative leads the chart, and where insufficient data is declared as such.
My own server, an encrypted tunnel, deployments with backup and rollback, and a three-step verification.
Language models, vector search and speech transcription on my own server, for data that cannot leave.
Third-party APIs, WhatsApp, Google Apps Script and business flows that stop depending on someone remembering.
Hypothesis testing, cohorts, retention and the discipline of not reporting a percentage that does not hold.
Installable apps that work offline, with notifications and a scanner, for people who work standing up.