Zone 00 · Arrival

Matthew Joseph Swezey

Machine Learning Engineer

I build machine learning systems the way you'd build an island: solid bedrock of data, well-chosen blocks of tooling, and structures that stay standing in production long after launch day.

Explore the island Visit the harbor
8+years in applied ML
40+models in production
12 TBdaily data handled
Matthew Joseph Swezey
Resident ML Engineer
Descend

Zone 01

Bedrock

The foundation everything else is built on: who I am, how I work, and what I care about when shipping models.

I'm Matthew, a Machine Learning Engineer who turns promising models into dependable, measurable production systems.

For more than eight years I've designed, trained and deployed ML services for logistics, fintech and industrial teams: forecasting engines that plan port operations, retrieval systems that answer questions across millions of documents, and real-time detectors that flag anomalies within milliseconds.

My work sits where modeling meets engineering. I care about clean data contracts, reproducible training pipelines, honest evaluation and monitoring that catches drift before users notice. I prefer small, measurable iterations, readable code, and models that earn their place by moving a real business metric.

Measure before you modelShip small, monitor everythingModels are products
Home tileKotka, Finland
SpecialtiesLLM & retrieval systems, forecasting, MLOps
CurrentlyOpen to senior ML engineering & consulting work

Zone 02

Toolblocks

Every block is a tool I build with. The taller it stands, the deeper the experience. Hover a block (or a name) to raise it.

Zone 03

The Island

Each building on the island is a project I've built. Tap a building, or pick a district, to step inside.

District 01 · Time-series forecasting

Harbor Forecast

A multi-horizon demand forecasting platform for container terminal operations. Temporal Fusion Transformers and gradient-boosted baselines predict vessel arrivals, crane load and yard occupancy up to 14 days ahead, feeding the daily staffing and berth planning tools.

−18%forecast MAPE
14 daysplanning horizon
2,300series nightly
PyTorch ForecastingLightGBMAirflowBigQuery

District 02 · LLM & retrieval

Lumen Search

Retrieval-augmented question answering over 3M+ technical documents. Hybrid BM25 and dense retrieval, a fine-tuned cross-encoder re-ranker and grounded LLM answers with citations, evaluated continuously against a curated golden set.

+31%answer accuracy
85 msp95 retrieval
3M+docs indexed
TransformersFAISSFastAPIRay Serve

District 03 · Computer vision

Canopy Vision

Semantic segmentation of drone imagery to map forest health and storm damage across large boreal plots. A U-Net with an EfficientNet encoder, trained with weak labels and active learning, turns days of manual surveying into an overnight batch job.

0.87mean IoU
40×faster surveys
60ktiles labelled
PyTorchsegmentation-modelsLabel StudioGDAL

District 04 · Real-time ML

Sentinel

Streaming anomaly detection for payment transactions. Sequence autoencoders and an isolation-forest ensemble score events from Kafka in real time, with SHAP-based explanations surfaced to analysts for every alert.

12 msscoring latency
−42%false positives
9k/sevents scored
KafkaPyTorchONNX RuntimeSHAP

District 05 · MLOps platform

Feature Foundry

An internal platform that gives every team the same path from notebook to production: a shared feature store, experiment tracking, automated evaluation gates and one-command deployment to Kubernetes with drift monitoring built in.

5 days → 4 htime to deploy
14teams onboarded
300+shared features
FeastMLflowKubernetesEvidently

Buildings rise when you hover them. The selected one stands tallest.

Zone 04

Terraces

A career built level by level. Each terrace adds a block to the stack, and the newest sits at the top.

Level 04 2023 – Present

Senior Machine Learning Engineer · Northwave Analytics, Helsinki (hybrid)

  • Lead a team of four engineers building LLM-powered search and document intelligence products.
  • Designed the evaluation harness and guardrails that took retrieval QA from prototype to 30k weekly users.
  • Cut GPU inference cost by 38% through quantization, batching and model distillation.
Level 03 2020 – 2023

Machine Learning Engineer · Baltic Freight Intelligence, Kotka

  • Built forecasting pipelines for port throughput and vessel ETAs used in daily operational planning.
  • Introduced MLflow, CI-tested training pipelines and automated model retraining on Kubernetes.
Level 02 2018 – 2020

Data Scientist · Kymi Process Systems

  • Developed predictive-maintenance models on sensor data from paper and pulp production lines.
  • Reduced unplanned downtime by 21% with early-warning models deployed at three plants.
Level 01 2016 – 2018

M.Sc. Computer Science · Machine Learning track

  • Thesis on uncertainty estimation for deep time-series forecasting models.
  • Teaching assistant for Introduction to Machine Learning and Data Engineering courses.

Zone 05

Harbor

Every island needs a harbor. Drop anchor here to get in touch about roles, collaborations or consulting.

Harbor Office

Messages are collected daily. I usually reply within one or two working days.

NameMatthew Joseph Swezey
Address
Saarelaistentie 8, 48310 Kotka, Finland
Send a message Call the harbor