Five different systems. One method.

The system

An ERP written in GeneXus in 2000, on an Oracle database with hundreds of millions of records, that seven departments depend on: sales, maintenance, warehouse, safety, logistics, HR and finance.

Diagnosis

A large share of daily work was copying data from one system to another. A single freight quote took 15 to 18 minutes, and the team prepares 20 to 30 of them a day.

Treatment

• Integration of the Oracle database into the new TypeScript ERP (microservices, Docker, CI/CD) without losing the business logic the old system carried.
• Demand forecasting models that combine the ERP with external market sources.
• Automation of freight quoting and, with the same approach, of international metal price tracking, order entry, electronic invoicing, HR and port-terminal integrations, and competitor price analysis.
• A company-specific MCP server that encodes business rules, architecture standards and mandatory controls, so the systems team can work with Claude Code without losing quality or traceability.

Outcome

Freight quotes went from 15–18 minutes to under one.

The system

The chain that follows every log from the standing tree to dispatch. It was kept by hand in spreadsheets and had to reconcile with the regulator’s public and private systems.

Diagnosis

I tested the rules the team took for granted against about 30,000 records, and several did not hold. The worst was a species key, already in production, that merged two species the regulator keeps apart. It affected 74% of incoming timber.

Treatment

• The rules for lot assignment, volume calculation and source attribution, written from how the operation actually works rather than from what was assumed.
• An ingestion engine that removes manual data entry at goods receipt: a streaming extractor on Cloudflare Workers for the regulator’s public portal, which has no API, plus an authenticated client for its operations book.
• AI reading of handwritten transport guides: several models with automatic failover, two parallel readings compared field by field, and human review for anything that cannot be verified. The validation arithmetic does not depend on the model.
• The traceability modules in React and TypeScript on Firebase: lot classification, plant coding, purchased-lumber registry and per-contract volumes.

Outcome

A traceability chain built on measured rules instead of assumed ones, and handwritten guides read with no errors on any verifiable field.

The system

The economy seen through daily prices, and how it relates to the official inflation index, U.S. Treasury yields, the dollar and Bitcoin.

Diagnosis

Nobody had systematically compared the company’s data with the official U.S. CPI. When I did, trend changes showed up about 45 days earlier on average.

Treatment

• Python pipelines that score institutional inflation forecasts against every official release.
• Models of the timing between inflation and different assets, and a hysteresis-based algorithm that classifies inflation regimes.
• An inflation-driven asset allocation strategy, validated through backtesting, which we adapted to on-chain infrastructure with the founders of QuantAMM.
• Truflation Argentina, with Argentina’s Ministry of Economy: a pipeline that ingests and AI-classifies the national price feed, published as open data.

Outcome

The flagship product stopped being pitched as another inflation measure and became a leading indicator. The QuantAMM strategy took first place at Chainlink Demo Day, SmartCon 2025, in New York.

The system

Four international diagnostic trials funded by the U.S. NIH and Abbott, with sites in Peru and Uganda. Today, also an AI cough-monitoring study with McGill, the University of Texas Medical Branch and Heidelberg.

Diagnosis

Data quality is not negotiable here: every record must trace back to its source, and any bias in a model is paid for in someone’s health.

Treatment

• Study databases with field-level validation, staged quality control and source data verification.
• The data architecture and analysis plan for the cough study: the longitudinal audio and clinical database that will train a neural network to recognise five respiratory infections from the sound of a cough.
• Evaluation by AUC, sensitivity and specificity, with algorithmic bias checks by sex and subgroup, and density maps across two Lima districts.

Outcome

Eight peer-reviewed publications, including one as first author in PLOS Global Public Health.

The system

A plant with more than a hundred cameras spread across recorders of different generations, operators working near suspended loads, and a safety team that cannot watch every screen at once.

Diagnosis

Before training anything I went through every camera and recorder. Many record at 960×480, where a glove is about 5 pixels wide and safety glasses 1 or 2. No model can see that, so the system only promises what the image allows and says which cameras are worth replacing. I also checked the starting training data by hand: it had real steel straps left unlabelled, and off-the-shelf models found 0 of 127.

Treatment

• A per-operator check of protective equipment that combines pose, silhouette and colour: helmet, ear protection and gloves, with a vote over time so a single frame never triggers an alert.
• A danger zone under the suspended load that follows the hook and compensates for camera movement, plus production measurements from video, such as the speed of an uncoiler.
• YOLO11 and RF-DETR compared on the same data. RF-DETR was more accurate and its licence allows selling it, so it powers the product.
• The full platform: a 3D plant view with anonymous people, incidents with WhatsApp alerts, compliance reports by area, and hardware sized for 30 cameras.

Outcome

A finished product that runs on real plant footage. It uses no facial recognition: it sees the risk, not the person, and blurs faces.