How a quant uses it
Eleven steps in the daily workflow of a quant at a $1B-AUM systematic equity fund, mapped to where pyquant.io helps and where it is neutral. Steps where the platform adds no value over existing tools are labelled as such, so you can see where the real bottleneck is.
| # | Today | With pyquant.io | Value |
|---|---|---|---|
| 1 | Open laptop, set up environment. Fight with conda/pip/poetry; Docker for some deps; hours-days. | One `make demo-fresh` command. Full local stack — Postgres, MinIO, mock OIDC, MLflow — in 5 minutes. | High |
| 2 | Connect to data. Per-source auth; per-source schema discovery. | Data is already typed, validated, point-in-time correct in the gold layer. Polars dataframe is a function call. | High |
| 3 | Sample for exploration — pandas + ad-hoc SQL. | Polars lazy; one query API across silver and gold. | Medium |
| 4 | Iterate feature code in a research script. Copy-paste between research and production; functions slowly diverge. | Research code imports the same `features/` module the production serving path uses. One source of truth. | High |
| 5 | Test feature on sample. Visual inspection; ad-hoc unit tests. | Pandera schema validation runs automatically. Integration tests against the local docker-compose stack. | Medium |
| 6 | Commit to git. | Commit to git (no platform addition). | Table-stakes |
| 7 | Kick off real training. Slurm / SageMaker / internal k8s; wait hours. | One-click submission to Cloud Run Jobs (CPU) or Vertex AI (GPU). Each fold tracked as an MLflow run; progress and results visible in the run-detail UI. | High |
| 8 | Review trained model across MLflow + internal dashboards + SHAP plots scattered across scripts. | Run-detail UI shows metrics, walk-forward evidence, PBO/DSR, factor attribution, baseline comparison — one screen. Visual lineage from bronze to model version. | High |
| 9 | Hand off to engineering for deployment. Multi-week back-and-forth; re-implementation; QA cycle. | NO HAND-OFF. Quant promotes in the registry UI; audit log records it; serving lazy-reloads. End-to-end in seconds. | Highest |
| 10 | Monitor in production — custom dashboards, usually inadequate; oncall is reactive. | Inference log built in. Per-inference drill-down. Alerts on rate / latency / error / drift. | High |
| 11 | Receive LP ODD questionnaire — scramble for evidence; 14-page doc; LP questions reveal gaps. | Share-screen the UI. LP asks "show me a specific inference 18 months ago" — appears in 60 seconds. LP asks "what was your walk-forward methodology" — the screen shows it. | Highest |
Why step 9 matters
Step 9 is where the platform earns its keep. Today, promoting a trained model to production is a multi-week back-and-forth between the quant who trained it and the engineer who re-implements the feature code in the serving pipeline. Every team rediscovers the same class of bug: the feature that worked in research produces different numbers in production.
With pyquant.io there is no hand-off. The Python function that computes a
feature at training time is the same function that runs at serving time — it travels with the
model artefact as an MLflow pyfunc
wrapper. The quant promotes the model in the registry UI. The audit log records the promotion. The
serving role lazy-reloads. End-to-end in seconds.