First 90 Days and Onboarding Plan Questions

Success in my first 90 days would mean: (A) I’m fully onboarded into the engineering workflow and trusted to make small architecture and model decisions, (B) I’ve delivered 2–3 measurable quick wins that improve a key pipeline metric, and (C) I’ve built relationships with core stakeholders.

30-day (Learn & Align)

• Meet stakeholders (priority order): direct manager, team lead/PM, infra/SRE, data engineering, MLOps, product/UX, research scientists. Weekly 1:1s to set expectations.
• Read & run: core repo READMEs, model training/eval pipelines, CI/CD, infra runbooks. Reproduce a baseline training/eval locally.
• Datasets: inventory data catalog, privacy/compliance docs, sample slices used in production.
• Secure access: dev GPU node, VPN, code repos, dataset buckets, experiment tracking (Weights & Biases / internal), build/test creds.

60-day (Contribute & Prototype)

• Own a small bug/feature in model pipeline; add tests and improve CI latency.
• Prototype a model tweak or data augmentation on a representative slice; track metrics.
• Work with MLOps to containerize a training job and schedule on infra.

90-day (Deliver & Handoff)

• Quick wins (measurable):
  1. Reduce end-to-end training time for a common experiment by 20% by optimizing data pipeline / batch sizing — measured by average epoch time.
  2. Improve a baseline model’s F1 or latency on a prioritized slice by 3–5% (or reduce inference latency by 10%) with targeted fine-tuning/quantization — measured in benchmark suite.
  3. Ship a reproducible notebook + automated test that documents a key experiment and decreases onboarding time for future hires by providing runnable baseline (measured by checklist completion / feedback).

I’ll document decisions, share a 90-day readout with metrics and next-step proposals, and propose a 6-month roadmap aligned to product goals.

30 days — Learn & stabilize

• Deliverables: inventory of datasets, models, infra; runbook for dev environment; reproduce one baseline model training end-to-end on dev GPU; stakeholder map.
• Success metrics: able to run training/evaluation pipeline in ≤2 hours; dataset access for core teams; list of top-5 model failure modes.
• People to meet: ML lead, Data Engineer, MLOps/SRE, Product Manager, 2 domain SMEs.
• Risks mitigated: hidden data access & compliance issues, missing reproducibility, environment drift.

60 days — Improve & prototype

• Deliverables: one improved prototype (e.g., fine-tune or small architecture change) showing measurable lift; automated data validation checks; CI training job configured; cost-performance baseline report.
• Success metrics: prototype improves key metric by ≥X% (or reduces latency by Y%), CI runs model tests on PRs, data quality alerts on ingestion.
• People to meet: QA, Security/Privacy, Cloud infra owner, 2 product owners for use-cases.
• Risks mitigated: model regression in prod, data drift, deployment failure modes, untracked costs.

90 days — Production-ready & cross-team enablement

• Deliverables: productionization plan + rollout checklist; archived experiment results + reproducible pipelines; monitoring dashboard (metrics, drift, latency); knowledge transfer workshop & docs.
• Success metrics: production readiness checklist complete, SLOs defined (accuracy, latency), monitoring detects injected drift in tests, two engineers trained to run pipelines.
• People to meet: Engineering manager, stakeholders from ops and business, legal for compliance sign-off.
• Risks mitigated: lack of observability, single points of knowledge, regulatory non-compliance, unanticipated scale issues.

Ongoing priorities: schedule biweekly syncs with stakeholders, maintain risk register, and plan quarterly model review cadence.