Data Scientist · Universidad Mayor · Santiago, Chile
This profile collects work and personal projects organized by the kind of problem they solve: forecasting, catching anomalies, anticipating failures, measuring causes, optimizing, managing risk, getting data in order, and working with text and images. The point is to show the different ways a data scientist can help an organization, in mining, energy, finance, retail or pensions.
The same rules apply throughout: the code runs end to end, the numbers come from that run, nearly every repo has tests and CI, and negative results are published like positive ones.
Four projects on real data with a finding that was not the expected one.
24 years of daily unit values, in UF. Switching to Fund E at the bottom of a crash costs 1.9 points of real return a year; switching earlier gets the same Sharpe as a static mix: there is no timing, there is less equity.
Did the maintenance program work, and on which trucks? On real sensor data from 60,000 Scania trucks, the default DRLearner collapses (r = −0.01); the cause was isolated and it recovers to 0.61–0.79.
3,000,888 sales rows. A quarter of the "zero demand" was stores that had not opened yet, and the model with the best metric leaves a stockout on 41% of store-days.
Checking an A/B test every day inflates the false-positive rate from 5% to 24.2%, and the Bayesian rule usually assumed "safe" barely helps (20.5%).
A Python library to download the UF, dollar, copper, pension fund unit values and CMF bank delinquency, clean and in a DataFrame. It fixes real source errors, like the dollar leaking into the UF in 2014 or the CMF Excel files' three layouts, and checks every week that the sources haven't changed.
import datoschile as dc
dc.indicadores.uf(desde=2020)
dc.pensiones.indice(fondos="A", desde=2008, real=True)
dc.cmf.morosidad(solo_sistema=True)Each of the 17 depots tracked fuel in its own Excel file. First I wrote a Python script that merges the sheets, computes each bus's km per litre against its previous load and flags the ones outside their model's range. Then an AppSheet app that replaced the sheets: operators record loads, pump readings and tanks from their phone, and the app computes per-pump consumption, theoretical stock and AdBlue. Reviewing the code, I fixed three bugs: dates sorted wrongly across months, buses left without a range because of how the standard was written, and every bus's first load flagged by mistake.
34 projects. Each states whether it uses real or simulated data.
Forecast scored as a purchase decision: quantiles only pay above a ~3:1 cost ratio.
Panel from 128 CMF Excel files. Only ARIMA at 3 months beats naive; adding unemployment, policy rate and IMACEC makes it worse.
Solar, wind, demand and marginal cost at 5 nodes. Solar: 3.64% WAPE vs 26.95% naive; on wind, naive wins.
TBATS: 2.47% MAPE vs 6.54% seasonal naive; ETS's edge only shows under rolling validation.
14 years of real LME prices: GARCH(1,1) has the best QLIKE and only HAR-X, with the VIX and the dollar, ties it. CatBoost re-tuned in every fold does not beat it.
24.8M Binance trades: at 30 s, persistence (RMSPE 4.58) beats LightGBM and the neural net.
ROC-AUC makes a model with 3.4× worse PR-AUC look almost equal (0.931 vs 0.965); label-free graph AML: ROC-AUC 0.893.
The statistical baseline lands below chance (ROC-AUC 0.383) and ensembles don't win. 266 tests.
13.7 M ChileCompra order lines: Compra Ágil orders pile up under the cap and the pile moved with the law, but there is little sign of split purchases. Isolation Forest finds 38.8% of planted anomalies reviewing 5%.
C + Ruby + Python, each layer profiled: 31 ns per call in C; 4.68 ms p50 end to end.
EWMA and CUSUM validated on the March 2020 crypto crash, at 7.26M ticks/s.
Zero false positives outside the injected shock across 43,200 windows.
23,550 trucks: −31% official cost on test, but the saving runs out if a visit costs double, because the model overstates risk.
Bearings (NASA IMS) and seismic signal (LANL): spectral features with boosting beat a CNN by 2.7×.
Kalman estimates ore hardness with 79% less error; 24 h energy forecast 27.6% better than Holt-Winters.
Uplift and staggered DiD; the ATT matches a reference implementation on a real panel.
Sample size, SRM, CUPED and multiple testing, with a harness that checks each rule's error rate.
A regime signal that "earned" 1.4–2.3% a year ties a static mix on Sharpe out of sample. In real terms, 2021–23 fell 26%.
Real iron-ore plant: R² 0.83 on a random split is -0.43 in time order; only recent lab results beat persistence, and four models prescribe opposite pH moves.
Multi-depot routing with time windows over real polygons: 0 of 173 zones unserved.
Threshold by customer value: +US$75,847 contacting 78.6%, vs +US$42,717 contacting everyone.
26 techniques: R+Python+C scorecard at 270.6M rows/s, IFRS 9 PD, bias audit and a shadow/canary/retraining lifecycle.
Six languages on Central Bank data. The LSTM (51.0%) doesn't beat the majority class (53.6%).
Eight techniques. All five strategies lose after costs; spoofing detection works (0.92 precision).
C++20 Monte Carlo under GBM, Schwartz and Heston: 9.64× on 16 threads and a real bias fixed.
One toolkit on Central Bank, COCHILCO and World Bank data; it caught a corrupted source value.
dbt + DuckDB from staging to marts, 83 quality tests and model-ready views.
Polars + DuckDB over Parquet; repeating the benchmark 7 times changed the conclusion.
Higgs boson (AMS 3.64 vs 3.8–3.9 for the winners) and Kepler exoplanets (79.3% vs 49.8%).
ML.NET + ONNX in a desktop app: P80 at R² 0.957.
Hybrid search over DS 132; re-ranking lifts MRR from 0.920 to 0.981.
An LLM that answers by calling tools (SQL, scoring, anomalies) instead of making up numbers.
Offline inference, guardrails and QLoRA; the base model makes up arguments in 94 of 100 prompts.
68% fewer parameters and ~10× faster on a Raspberry Pi, at mAP50 0.175 vs 0.212.