GITHUB COPILOT / DATA SCIENCE
Français Download PowerPoint Download PDF

Data science / VS Code

GitHub Copilot
for data scientists

Less boilerplate.
More time for scientific judgment.

Complete code. Ask better questions.
Delegate bounded work, then verify.

A practical briefing for Desjardins teams

one workspace / three scientific habits

  1. Explore

    Understand data and assumptions.

  2. Experiment

    Build and compare reproducible baselines.

  3. Validate

    Test code, metrics and leakage risks.

02 / Choose your interaction

Three ways to work with Copilot

Choose the interaction by how much work you want to delegate.

COMPLETION

Keep typing

Inline suggestions

Finish a pandas transform, a SQL query or a test assertion.

You steer every edit.

# aggregate transactions
# by customer and month
ASK

Think it through

Conversational help

Explain a notebook, compare metrics or challenge a split.

Advice first; apply deliberately.

"Could this feature leak future information?"
AGENTIC

Delegate a task

Agent + tools

Edit related files, run tests and iterate on a bounded task.

Review diffs and tool actions.

"Build a baseline pipeline and run its unit tests."

More autonomy does not mean less accountability. Keep context scoped and permissions intentional.

Completion is an editor surface, not a chat mode. UI labels vary by version and harness. Inline suggestions / Chat in VS Code

03 / Evidence before edits

RPI: turn a question into verified work

HVE Core on GitHub: Research → Plan → Implement, with review and follow-up.

Research

Understand before editing

  • Inspect data contracts and code.
  • Check leakage and alternatives.
  • Separate evidence from assumptions.

Output: findings + open decisions

/rpi-research

Plan

Make success testable

  • Choose the split and metrics.
  • Sequence small, bounded tasks.
  • Define tests and acceptance criteria.

Output: plan + critique

/rpi-plan

Implement

Execute the approved scope

  • Build the pipeline and tests.
  • Run checks; record real results.
  • Review diffs against the plan.

Output: changes + validation

/rpi-implement
Reuse adequate researchReview the planValidate resultsRoute gaps to the right phase

HVE is an optional framework, not a default VS Code feature. Install/configure its RPI Agent or phase skills. microsoft/hve-core / RPI lifecycle and commands

04 / Spend reasoning where it matters

Right-size the model to the uncertainty

Start with routing, or choose capability manually. Escalate when evidence says you should.

LARGE

Research + rubber ducking

Challenge assumptions, compare approaches, look for leakage and counterexamples.

MEDIUM

Planning

Turn evidence into tasks, dependencies and measurable acceptance criteria.

SMALL / MEDIUM

Implementation

Small for routine edits; medium for complex changes. Escalate if failures or ambiguity persist.

*Hydrafusion is setup-dependent, not a universal Copilot option. Capability tiers are a practical heuristic, not product labels or reasoning-effort settings.

Rubber ducking means asking the assistant to challenge your reasoning. A convincing response still needs verification. Auto selection / Model availability

05 / A bounded experiment

Try it: a reproducible credit-risk baseline

One synthetic dataset. One clear objective. Three bounded prompts.

RESEARCH / ASK
Inspect this synthetic dataset and notebook. Identify leakage risks and compare validation splits. Do not edit files yet.
PLAN / HVE RPI
Plan a reproducible baseline with a time-based split. Fit preprocessing on training data only. Define metrics and tests before coding.
IMPLEMENT / AGENT
Implement the approved plan and run the tests. Report PR-AUC, calibration and subgroup results. Show the diff and any remaining limitations.

Banking guardrail: synthetic or approved de-identified data only; no client PII or secrets in prompts.

Presenter notes and sources

This is an educational workflow, not a production credit-model approval. Choose thresholds with the business before implementation. Account for label horizons, customer isolation and temporal gaps when the data requires them. Reserve the final holdout for final evaluation.

PR-AUC is not a complete decision criterion. Calibration, uncertainty and subgroup analysis need context; subgroup results do not establish fairness or compliance. A clean rerun uses documented dependencies and seeds and a fresh kernel/environment.

Customer policy governs approved data, providers, tool permissions and retention. De-identification requires approval and does not guarantee privacy. The large/medium/small mapping is a workshop recommendation, not an HVE requirement. Hydrafusion behavior and availability were not verified in public Copilot documentation.

Sources checked October 20, 2026. Product availability and labels can change.