Lab log
Gaspar Quintana
Data science, machine learning and AI automation. I write up every project as an experiment: the question, the method, the result and what I learned. When a result doesn't hold up, it stays in the log too.
status: open to data roles
- entries
- 17
- retracted
- 2
- EXP-01 · 70.7% → 91.7%
Córdoba, Argentina
§ 01
Experiments
Six selected entries. Each one opens with the question I wanted to answer.
EXP-01computer vision
Reading land use from space
- question
- Can a CNN trained from scratch classify 64×64 Sentinel-2 tiles into 10 land-cover classes?
- learned
- More data beat every architecture change: +7.3 points from 500 to 1,000 images per class.
test accuracy · 91.3% on 17,000 never-seen images
Open entry →
EXP-02reinforcement learning
An agent that crosses a minefield
- question
- Can a DQN agent learn a 7×7 minefield with tunnels, starting from any cell, with no map?
- learned
- A −1 reward per step taught it the shortest safe path: 14.7 → 5.4 steps.
success over 1,000 games · random player vs. trained DQN · 0 bombs hit
Open entry →
EXP-03deep learning
Does it read my handwriting?
- question
- Does a 99% MNIST model still work on digits I drew myself, outside the dataset?
- learned
- Even with 200 of 784 pixels it keeps 97.4%: most pixels are noise.
test accuracy (95% CI 99.24–99.54%) · my own digits recognized
Open entry →
EXP-04AI automation
A bot that survives its sources
- question
- Can a Telegram bot keep recommending events when the websites it scrapes break?
- learned
- Design for degradation: fall back, log every failure, alert a human.
n8n · Gemini agent · voice via Groq · logs, alerts and /metrics
Open entry →
EXP-05BI · forecasting
Next week's profit, honestly
- question
- Can a small business forecast next week's profit from 4 years of messy spreadsheets?
- learned
- Check what the target really is: 'future' profit was same-week profit.
R² after removing the leak (synthetic data) · 33% less error than naive
Open entry →
EXP-06client project
A barbershop goes online
- question
- Can a barbershop move off WhatsApp and AgendaPro without losing its customers?
- learned
- Clean the data before you automate: 1 in 6 records was a duplicate.
records → real customers · 112 duplicates merged · 5 automations
Open entry →
ALL ENTRIES
The rest of the log: foundations, statistics, BI and early analyses.
- 2026Heart disease prediction with ensembles and cross-validationAn L1-regularized logistic regression scored best (83.8% 20-fold cross-validated accuracy), ahead of every ensemble. The best ensemble was the soft VotingClassifier (82.2%), followed by Extra Trees (79.5%) and Random Forest (77.1%).ML ↗
- 2026Statistical inference with chi-square, ANOVA and mutual informationHaving dependents (χ² = 189.1) and a partner (χ² = 158.7) are associated with churn, while gender is not (p = 0.49). ANOVA was applied only to the one variable that met its assumptions.Statistics ↗
- 2026Computer vision fundamentals with OpenCVA hand-built smoothing and edge-detection pipeline, the groundwork for the EuroSAT CNN project.Vision ↗
- 2026Natural-language search engine for computer peripheralsA working API that answers free-form queries over the full peripherals catalog.NLP ↗
- 2025Neural network foundations, from the perceptron to a tuned DNNThe regression DNN reached a test RMSE of 89.8 (5-fold CV, 92.6 ± 40.7), with fold-to-fold variability documented as a model limitation.Deep learning ↗
- 2025Breast cancer diagnosis with DNNs, SVMs and PCAThe L2-regularized SVM reached 97.6% cross-validated accuracy and 94.7% on test (F1 0.947), the DNN 94.5% on test, and 3 principal components keep 81.9% of the variance.Deep learning ↗
- 2025SQL Server sales dashboard with monthly forecastingA dashboard regenerated by a single procedure, plus an on-demand monthly forecast.SQL ↗
- 2024Titanic, feature engineering and Random ForestWithout data leakage, the Random Forest scores 81.6% on the hold-out (84.8% cross-validated), 3.9 points above the "women survive" rule. The original 98.9% came from computing family and ticket survival rates on the whole dataset.ML ↗
- 2024Buenos Aires real estate, exploratory analysisFor the 3 most-listed property types in Buenos Aires City (91,485 listings), rooms and bedrooms are the best individual price predictors, with moderate correlation.Analysis ↗
- 2024Blood pressure by age, cleaning and analysisHigh blood pressure goes from about half of cases at ages 18–29 to nearly all cases after 70, with the sharpest jump between 40 and 59.Analysis ↗
- 2024Power BI dashboards for IT support and salesAgent performance, satisfaction and sales KPI panels, filterable by year, type and agent.BI ↗
§ 02
Retracted results
Two of my own results that didn't survive a second look. I keep them visible on purpose: finding the mistake is worth more than the pretty number.
RETRACTED
Titanic survival
what was wrong: Family and ticket survival rates were computed with the test passengers' outcomes. Without the leak, the model still beats the “women survive” rule by 3.9 points.
RETRACTED
Deloto weekly profit
what was wrong: The target, called “future” profit, was effectively same-week profit, and an unshifted rolling mean included it. Rebuilt to predict the following week (on synthetic data).
§ 03
In progress
EXP-07 · thesis
Unbiased valuation ranges
People who value companies type 5-year assumptions into a spreadsheet by hand: sales growth, margins and cash flow. The verdict depends almost entirely on those numbers, and the evidence says hand-made projections tend to be optimistic.
I'm building a model that gives a calibrated reference range for each assumption, based on what thousands of comparable companies actually achieved. It uses SEC filings as they were available at each date, quantile models with conformal calibration and walk-forward validation, and it is compared against baselines from the literature: random walk, mean reversion and base rates.
Preliminary results
§ 04
Log
Before data I sold cars and houses and worked at a newspaper. That's why I build tools with the people who will use them in mind.
Sep 2026 – now
Scalo · systems and automation for SMBs
AI automation and CRM
I design and maintain GoHighLevel automations for 9 businesses in healthcare, aesthetics, financial education and services: WhatsApp bots that triage inquiries, segmented campaigns with follow-ups, clinic schedules and reminders, contact deduplication and migration, integrations with external systems and scheduled jobs on Vercel. Every system ships with video walkthroughs.
2025 – now
Agua Canta · event venue
Stock, inventory and automation
A Telegram bot built with Claude Code that logs stock in and out in real time. Post-event closing dropped from 3 hours to under 20 minutes, with 500+ items tracked.
2024 – 2025
Deloto · Esencia Natural
Data analyst and automation
An n8n pipeline that cut the weekly report from 4 hours to 30 minutes, Power BI dashboards and SQL queries to validate data.
2023 – 2024
Transformar · prefab houses
Sales advisor
End-to-end sales of a high-ticket product, with custom quotes.
2022 – 2023
Alpha Motors
Sales advisor and team lead
40+ inquiries a week and error-free paperwork for every deal.
2020 – 2022
Galpón Gráfico MX · El Milenio · La Dupla
Content, journalism and production
Stories for print, radio and web, video editing and quality control in production.
EDUCATION
Associate degree in Artificial Intelligence and Data Science
Universidad Siglo 21 · Colegio Universitario IES
24 of 24 courses passed · thesis in progress
AWS Certified Cloud Practitioner
Amazon Web Services
2025
Technology and computing · mathematics and computer science
UTN · FaMAF (UNC)
courses completed
Spanish (native) · English (B2/C1) · basic Italian and Portuguese
INSTRUMENTS
- modeling
- Python · scikit-learn · TensorFlow/Keras · PyTorch · LightGBM · Stable-Baselines3
- data
- pandas · NumPy · SQL Server · PostgreSQL · Power BI · Excel
- automation
- n8n · Claude Code · FastAPI · Telegram bots · LLM agents
- tooling
- Git · Docker · Google Colab · AWS
§ 05
Let's talk
I'm looking for a full-time data analyst or data scientist role. I'm based in Unquillo, near Córdoba, Argentina, and available to start right away.