7 years building systems, 6+ years building models.

01

Unusual, given the drivers

Bayesian anomaly detection on demand signals

• I built a service that asks of demand what I first asked of train telemetry: is this point out of line with what I expected? A rolling Normal-Inverse-Gamma regression on promotion, price, seasonality and weather gives a Student-t posterior predictive at each step. A two-sided posterior predictive p-value flags the tail.
• The expectation moves with the drivers. A promotion week expects more demand, so only the part the promotion does not explain gets flagged. The service runs as a FastAPI container on GCP Cloud Run, and I tagged it v1 on 8 September 2025 with the method written up on Confluence.

Basis Build and deployment facts from my journal, August to October 2025, tested on generated demand signals. I marked the detection evaluation done on 15 October 2025 without recording a figure, so no detection rate is quoted.

deployed as a serviceon GCP Cloud Run, in Avathon's demand planning platform

Python, Normal-Inverse-Gamma updating, scikit-learn BayesianRidge, PyMC, FastAPI, Docker, GCP Cloud Run

02

What the rules never fired on

Unsupervised anomaly detection on train telemetry

• I developed unsupervised anomaly detection on train odometry and ATC subsystem signals using autoencoder architectures. It reached over 99% recall and surfaced failures that rule-based monitoring missed.
• A statistical fault detection model that I packaged and deployed into the fleet management system improved maintenance efficiency by 30%.

Basis Measured by the Alstom team, 2020 to 2023.

in productionon a fleet management system

Python, PyTorch, TensorFlow, autoencoders

Show the working

03

Which driver moved demand

Attributing a flagged anomaly inside a multi-agent planning co-pilot

• At Alstom the question was which fault caused a failure. Here it is which driver caused a flagged anomaly. I drop each candidate in turn (promotion, price, weather, competitor stockout), refit on the anomaly window and rank them by ΔBIC or Bayes factor. The posterior coefficient and its credible interval size each effect.
• Between 29 September and 1 October 2025 I built a planning co-pilot around it. A root agent passes a planner's request to anomaly, causal and forecasting agents through structured tool calls, under an Observe Orient Decide Act loop, and GPT-4o returns the recommendation with its explanation.

Basis Method from my journal, September 2025; the four agents built 29 September to 1 October 2025. Tested on generated demand in which a weather shock hit only some states. The attribution evaluation was still open on 15 October 2025, so no accuracy is quoted.

presented internallythe attribution method, at an Avathon monthly all-hands

Python, scikit-learn BayesianRidge, ΔBIC, Bayes factors, LangChain, GPT-4o, structured tool calls

04

A ranked cause, not a correlation

Causal Bayesian networks on urban transit failures

• I built a root cause system using causal discovery and a Causal Bayesian Network. It answers interventional and counterfactual queries over Bayesian posteriors, so engineers get causes ranked by probability instead of correlations.
• Troubleshooting time fell 86%, and the method became a UAI 2024 paper on time-dependent counterfactual root cause analysis.

Basis Measured by the Alstom team, 2023 to 2025.

in productionon an urban transit fleet

Python, causal discovery, Causal Bayesian Networks, do-calculus, counterfactual queries

Show the workingRead the paper

05

Select on the holdout

What the fit-time winner could not tell me

• On 69 of 76 series (90.8%), the model that won at fit time was not the model that won on the holdout, so I retired the in-sample criterion.
• I rebuilt the criterion as WAPE primary, gated by event window WAPE at or below 1.2x WAPE and capped at 1.5x per-series degradation. Median holdout WAPE went from 0.574 to 0.451 across three pipeline reruns.

Basis Median across 76 weekly retail series, holdout window, single split.

in productionfor a global consumer-devices manufacturer

Python, LightGBM, XGBoost quantile objective, statsmodels ETS and SARIMAX, pandas, parquet

Show the working

06

Products that inherit demand

Lending a new product the history of the one it replaces

• I mapped each PC product family to the models it replaces, including many-to-many transitions, and fed the predecessor in as a feature beside product age and days left in the lifecycle. A new model with little history borrows signal from the old one. A family with a single record and no predecessor on file has nothing to learn from and nothing to borrow, so when I reviewed results across all families I marked it as not forecastable from history.
• The same question comes up wherever new products replace old ones. At a global consumer-devices manufacturer, one launch series overshot about 12 times over (WAPE 11.76) and another was forecast at zero because of pre-release clipping. I concluded that capped and launch series should be scored by cap correctness and launch-ramp shape, not by WAPE.

Basis Multi-echelon forecast of 80 SKUs over 5 regions, 2025, a scale fact: no accuracy was recorded for the transition features. Launch series: holdout WAPE, 2026.

deliveredto a global PC manufacturer

Python, pandas, XGBoost, LightGBM, SARIMAX, Prophet, PyTorch LSTM

07

Forty thousand materials, four demand classes

Routing a spares catalogue by its time-series properties

• I classified 40,447 maintenance and spares materials by Syntetos-Boylan demand type, with quantity share smooth 34%, lumpy 22%, intermittent 21% and erratic 18%. I flagged the 2,888 of 40,447 (7.1%) that sit on a class boundary as unstable rather than settled.
• AutoETS was collapsing to a level model, so I added a forced seasonal AutoETS entry to the model menu. It took one material from MASE 0.77 to 0.51 and another from 1.65 to 1.31.

Basis Per-material panel, one site, holdout MASE per series.

shipped to customera mining and industrial operator, where production model selection reads it

Python, statsforecast with AutoETS including forced seasonal ZZA, AutoTheta, Croston-SBA and TSB, conformal prediction intervals, LightGBM, Plotly

08

Fill rate is the verdict

Judging a forecast by the inventory it produces

• I moved evaluation from forecast error to inventory outcome and shipped the verdict: deploy on 9 smooth series, where test fill rate was 97.7% against the incumbent policy's 74.5% at lower inventory cost, and keep the incumbent on intermittent and lumpy series.
• A pre-deployment guardrail caught a 54 percentage point fill rate regression on one material before rollout. I traced it to a reorder point that had collapsed from 3,618 to 489.

Basis Test window 2026-H1, nested train, validation and test backtest with leakage tripwires, 57 tests.

decision deliveredto a mining and industrial operator

Python, statsforecast, LightGBM, NumPy and SciPy lognormal fitting, pytest

Show the working

09

What to build this week, and how to ship it

One program for the production plan and the freight plan

• I built a two-stage plan. A demand forecast by SKU, channel, geography and week feeds a mixed integer linear program that plans the Master Production Schedule and the pack-out jointly, with one colour per production week and weeks-of-supply targets of 12 for channels 1 and 2 and 13 for channel 3. Each week it picks air, fast boat or ocean under a 7-week ocean lead time and 14,409 units a week of pre-build capacity.
• Demand averaged 13,033 units a week against that cap (90.5%), but 23 of 91 weeks ran over it, peaking at 80,059 in the Presidents Day week, so when to build ahead is the decision. Replaying 52 weeks of actual demand, the plan cost $508K in freight against $1.53M all air and $763K all fast boat. That equals all ocean: with the pre-build scheduled, ocean could carry the year.

Basis 52-week backcast replay of actual demand, channel 3 (about 95% of sales) in one region, run 21 April 2026.

proof of conceptbackcast replay for a global consumer-devices manufacturer

Python, PuLP, CBC, pandas, XGBoost, SARIMAX

10

The leak I found in my own benchmark

Feature attribution against a published result

• A feature importance export showed a lookahead aggregate at 68.8% of model gain against lag_1 at 0.08%. That made the model a scale estimator rather than a model of demand dynamics.
• I measured the leak at 9.9 and 6.2 WMAPE points on intermittent and lumpy demand, retracted the published claim in writing, and recomputed all 16 sweep configurations without the leak. The ranking held.

Basis Intermittent and lumpy quadrants, walk-forward backtest, one site.

shipped to customercorrected findings reissued

Python, LightGBM, feature attribution, pandas