Inferpathio exists because model monitoring is still mostly manual — cron jobs, Slack messages, and spreadsheets. We built the lifecycle layer that closes the loop automatically.
In 2021, Kevin Nakamura was the ML platform lead at a mid-size financial technology company in Seattle, running a team responsible for roughly 40 production models spanning fraud detection, credit scoring, and customer churn. Those models worked. They just didn't stay working.
Every quarter, someone on the team discovered — usually because a downstream business metric had gone sideways — that a model's input distribution had shifted weeks earlier. The tooling gap was obvious: there was no systematic way to watch 40 models simultaneously, no automated trigger to retrain when drift was detected, and no immutable record of which model version had been running when an incident happened. The team tracked model states in a shared spreadsheet. Retraining was triggered by Slack messages and gut feeling.
Kevin spent two years trying to stitch together MLflow for versioning, Prometheus for metric scraping, and custom Python scripts for drift scoring. It worked — barely, and only as long as the team that built it stayed employed there.
In early 2023, Kevin moved to Seattle's Capitol Hill neighborhood and started building Inferpathio with two engineers from his former team: a full lifecycle platform that handles experiment tracking, model versioning, drift monitoring, and retraining automation as a connected graph rather than four separate tools.
Inferpathio is not a monitoring dashboard. It's not an experiment tracker. It's the connective layer between those things — so when drift is detected, retraining starts automatically, and the new model version is registered with full provenance back to the original training run. The loop closes without a human in the middle.
We're a small, bootstrapped team. We take on customers we can actually serve, not ones we need to fake references for.
We build Inferpathio so that call doesn't happen to you.
Former ML platform lead at a financial technology company in Seattle, where he managed 40+ production models on a team of five. MS Computer Science, University of Washington. Believes model monitoring should be boring — which means it needs to be automated.
Built drift detection systems at two healthcare analytics companies before joining Kevin as a co-founder. Open source contributor with merged PRs in scikit-learn's statistical test utilities. Knows more about Population Stability Index than is strictly healthy for a person.
Seven years in distributed systems engineering at a cloud database company before joining Inferpathio. Designed the content-addressed artifact storage layer and the event graph that tracks lineage across training runs. Runs on cold brew and Postgres query plans.
Every feature ships as a Python API first. If the SDK is clunky, the dashboard doesn't matter. ML engineers live in code, not browser tabs.
Inferpathio is not a framework. It wraps your existing training code without restructuring it. Your PyTorch loop, your XGBoost pipeline — unchanged. We observe; we don't impose.
We'd rather ship PSI and KS-test with well-documented thresholds than a dozen drift signals that fire false positives and erode trust. One reliable alert beats ten noisy ones.
Inferpathio is not a generic ML platform. We don't build model serving, feature stores, or training infrastructure. Every feature we ship is part of the monitor-detect-retrain loop — nothing else.
Start tracking experiments and monitoring drift in 15 minutes.