Start Here

The two roles, in plain English

This guide is calibrated against two open Data Scientist roles that live on the same spectrum but at different points:

• SentiLink — Staff Data Scientist, Full Stack ($220k–$260k + equity, US remote, 6+ yrs with PhD or 8+ yrs with Masters). Identity-verification & fraud platform — real-time APIs verifying hundreds of millions of identities. Build fraud models end-to-end: data acquisition → featurization → labeling → training → experimentation → production → monitoring. Python 3, PostgreSQL, AWS (EC2, S3, RDS, Redshift). The JD explicitly says "unusual insights drive our competitive advantage rather than optimization of new machine learning methodologies" — meaning: deep domain knowledge and inventive features beat sklearn novelty.
• Archetype AI — Data Scientist (San Mateo, on-site, 4+ yrs). Multimodal sensor-fusion platform — "Newton," a real-time multimodal LLM for physical-world AI. The role sits in GTM / Solutions Engineering: review customer assets (video, sensor streams), prepare time-series datasets, design prompts and n-shot examples, configure lens parameters, support Solutions Architects through customer evaluations. Python, Jupyter, signal processing, Encord-style labeling.

The SentiLink role is "staff IC who owns a model domain end-to-end." The Archetype role is "mid-level applied scientist who turns customer data into a working POC." Different seniorities, both full-stack applied: both expect you to go from raw data to a shipped artifact without a separate ML engineer to hand off to.

What the rounds typically test

Loops for full-stack / applied DS roles usually include:

1. ML fundamentals — supervised models for tabular data, evaluation metrics (especially under class imbalance), calibration, model selection. Expect "given these tradeoffs, which model?" framing, not whiteboarding gradient descent.
2. Feature engineering deep-dive — almost guaranteed at SentiLink. Often a take-home or live session: "here's a fraud-ish dataset, what features do you build and why?" Domain-flavored at every step.
3. Production / code quality — write production-ready Python with tests. SentiLink explicitly says "production code that can be relied on for real-time decision making." Archetype implies it via "executable iterative preprocessing cycles."
4. Applied design — "Customer brings X video stream / fraud signal — walk us through how you'd build a model / prompt / pipeline." End-to-end ownership signal.
5. Coding — Python data manipulation, light algorithms, occasionally a SQL screen.
6. Behavioral — "tell me about a project you owned end-to-end," "describe a time domain knowledge changed your modeling approach."

This guide covers all six.

The folder, in reading order

The numbering follows the order you should read in. Five sections:

Section A — Orient (read first)

Section B — Core DS concepts (the technical core)

Section C — Coding (DSA)

Section D — Production / Cloud

Section E — Reference & Execution

Study schedule

If you have 7+ days

• Day 1: 01, 02 (orient) → 03 (ML fundamentals)
• Day 2: 04 (features), 05 (fraud/imbalanced — for SentiLink)
• Day 3: 06 (prompts — for Archetype), 07 (time-series & signals)
• Day 4: 08 (pipelines), 09 (production ML)
• Day 5: 10, 11 (coding drills on a timer)
• Day 6: 12, 13 (AWS, MLOps), 14 (domain context)
• Day 7: Drill 15. Reread 16. Sleep.

If you have 2–3 days

01, 02, 03 (ML), 04 (features), role-relevant 05 or 06, 09 (production ML), 11 (drill 4–5 problems), 15 (drill), 16. Skim everything else.

If you have < 24 hours

01, 02, 04 (features — central to both roles), 05 or 06 (whichever applies), 09 (production ML), 15 (drill all questions), 16. Skim 03, 07, 13 headings only.

The single most important reframe

What this means tactically:

• When you get a modeling prompt, your first move is domain questions, not algorithm questions. "What does fraud actually look like in this data? Who labels it? How fresh are the labels? What's the cost of a false positive vs false negative for the customer?" That's how you signal you'll thrive here.
• When you describe past work, lead with features and data decisions, not models. "I noticed that X co-occurred with Y in 90% of fraud cases — built a feature for that — moved AUC from 0.82 to 0.88." That's the SentiLink/Archetype version of impact.
• When you discuss production, treat it as table stakes. Tests, monitoring, drift detection, retraining cadence — these are part of the deliverable, not the SRE's problem.
• When you get a vague open-ended prompt ("here's a sensor stream, customer wants to detect X"), don't reach for fancy models. Reach for iteration loops — what's the smallest preprocessing + model + eval you can ship today, and how would you improve it tomorrow?

What winning looks like

You don't need to be the deepest researcher in the loop. You need to be the candidate who:

1. Picks the right model for the job in under 30 seconds, and can articulate why (calibration vs ranking, latency budget, label availability, drift profile).
2. Builds features that encode domain insight — and can talk about how they discovered the insight, not just the feature.
3. Handles class imbalance / delayed labels / drifting distributions without panicking, because they've thought about these failure modes before.
4. Writes clean Python with tests, types, and small functions — like an engineer would, not like a notebook would.
5. Reasons about production from day one — latency, monitoring, retraining triggers, on-call.
6. Asks the right domain questions early, and uses the answers to shape the modeling approach.

If you can do those six things on demand, you're in.