Demand & revenue forecasting
Hierarchical forecasts at SKU × location × week, with prediction intervals and feature attributions ops actually understand.
Prophet · LightGBM · DeepARAI / ML · 05 of 05
Applied ML that ships into your operations.
Forecasting, anomaly detection, recommendation, decisioning — the unglamorous AI that compounds. Models that ship into your operations dashboard, your ERP, your warehouse — not a Jupyter notebook the analytics team forgot.
3.2×Avg. ROI · year 1
92%Models still live after 12mo
6 wkPOC → operations dashboard
Anomaly & drift detection
Real-time anomaly streams on metrics, transactions and operational data — with explainable root-cause suggestions.
Isolation Forest · PyOD · ProphetRecommendation systems
Two-tower retrieval + reranking. Cold-start strategies that don’t embarrass you on day one. Online A/B from week one.
two-tower · rerank · A/BDecisioning & uplift
Predict who’ll respond to which intervention — for pricing, retention, collections. Treat ML as a decision engine, not a dashboard.
uplift · CATE · MABComputer vision in ops
QA on the line, shelf compliance, document layout, defect classification. Edge-deployable, retrainable from labeled stills.
YOLO · ViT · ONNXNLP & document intelligence
Extraction, classification and routing across contracts, invoices, claims. Layout-aware, confidence-scored, human-in-the-loop.
LayoutLM · DocAI · HITLHow it works
An ML lifecycle that ships and stays shipped.
Most models die in the gap between training and operations. We engineer that gap shut.
The cycle is the product. Models are artifacts of the cycle.
Production ML isn’t a notebook — it’s a loop that runs forever. Every step is automated, instrumented, and owned by someone who can be paged.
- 01
Frame the decision, not the model
Who acts on the output, by when, with what alternative? If we can't answer that, we don't train.
- 02
Features as a product
A feature store with lineage, point-in-time correctness, and the same features in training and serving.
- 03
Ship into the workflow
The output lands in the tool people already use — ERP, ticketing, dispatch — not a side-car dashboard.
- 04
Operate like infrastructure
Drift detection, on-call rotation, retraining cadence, business-outcome dashboard — reviewed monthly.
Tech stack
Boring tools, reliable outcomes.
Classical ML still pays the rent. We use it without apology, alongside deep learning where it earns its keep.
Classical ML
scikit-learnXGBoostLightGBMCatBooststatsmodels
Forecasting
ProphetDeepARN-BEATSTFTARIMA / ETS
Deep learning
PyTorchTensorFlowHuggingFaceYOLOViT
Anomaly & uplift
PyODIsolation ForestEconMLCausalML
MLOps
MLflowKubeflowDVCMetaflowAirflow
Feature store
FeastTectonHopsworks
Serving & ops
SageMakerVertex AIDatabricksBentoMLRay Serve
Monitoring
EvidentlyArizeWhyLabsGrafana
From vision to victory
From use case to P&L impact, in six weeks.
A pragmatic ML rollout designed around the people who’ll act on the output.
01
Week 1Frame the decision
Who acts, by when, with what alternative. Success metric in dollars or hours, not AUC.
02
Week 2Data & features
Pipeline, leakage check, point-in-time joins. Feature store stood up from day one.
03
Week 3–4Train & validate
Cross-validate against business KPI, not just F1. Calibrate probabilities.
04
Week 5–6Ship into the workflow
Output lands in the existing tool. A/B vs. control. Outcome dashboard live.
05
OngoingOperate & retrain
Drift watch, monthly business-outcome review, scheduled retrains.
Where it earns its keep
Three use cases that compounded.
No splashy launch, no podcast tour — just a model that’s been in production long enough to be boring, and profitable.
Inventory that stopped guessing.
Hierarchical forecast across 8.4k SKUs × 140 stores. Replaced a spreadsheet + intuition workflow. Now drives the weekly replenishment plan.
−31%Stockouts
−18%Working capital
LightGBMProphetFeastDatabricks
Collections that actually collect.
Uplift model picks the right intervention (SMS, call, lawyer) per delinquent account. Beats rules + intuition, every cohort.
+22%Recoveries
−14%Cost-to-collect
EconMLXGBoostMLflow
QA without a QA team.
YOLO + ViT classifier on the line at a packaging plant. Catches mis-labels, deformed seals, missing caps in real time.
F1 0.94Defect class
3 linesLive
YOLOONNXJetson
Why ETY
Models that stay live.
92%Of models we shipped are still serving production after 12 months.
3.2×Median ROI in year one across applied-ML deployments.
40+Production ML features shipped across retail, fintech, ops & logistics.
MonthlyBusiness-outcome review on every operated system — not just model metrics.