Microsoft: Applied Scientist & Data Scientist

• Clarify: the input (topic plus optional documents), the output (PPTX/JSON slides with title, bullets, images, notes), enterprise data access, latency budget (under 30 s), and brand templates.
• Pipeline: 1. Intent and outline: the LLM produces a structured outline (sections, slide count) as JSON. 2. Retrieval per section: hybrid search over the user's permitted documents (SharePoint/Graph) → re-rank → top passages. 3. Slide generation: the LLM fills a slide schema (title, 3–5 bullets, speaker notes, citation IDs, a suggested visual) per section, with sections run in parallel. 4. Rendering: map the JSON to the template (python-pptx / Office.js), insert charts or images, keep the brand theme. 5. Review loop: the user edits, and "regenerate slide 4" re-runs only that slide.
• Quality:
• Faithfulness: every bullet must cite a source.
• No overflow (bullet length limits).
• Human ratings.
• A rubric for an LLM judge.
• Safety and scale: access control respected at retrieval, PII filters, caching of document embeddings, cost control (a smaller model for the outline, a larger one for content).

• The loss for one example is ℓ(w) = −y log σ(z) − (1−y) log(1−σ(z)), with z = wᵀx.
• Gradient: ∇ℓ = (σ(z) − y) x.
• Hessian: ∇²ℓ = σ(z)(1 − σ(z)) x xᵀ.
• σ(z)(1−σ(z)) ≥ 0 and x xᵀ is positive semi-definite, since vᵀ(x xᵀ)v = (xᵀv)² ≥ 0. So the Hessian is PSD.
• The sum over examples (plus an optional L2 term λ‖w‖², whose Hessian is 2λI) stays PSD, so the loss is convex, and with L2 strictly convex. Any local minimum is global.

import numpy as np, matplotlib.pyplot as plt

def pr_curve(y, s):
    order = np.argsort(-s); y = np.asarray(y)[order]
    tp = np.cumsum(y); fp = np.cumsum(1 - y)
    precision = tp / (tp + fp); recall = tp / y.sum()
    return np.r_[1, precision], np.r_[0, recall]

y = np.array([1, 0, 1, 1, 0, 0, 1, 0]); s = np.array([.9, .8, .7, .6, .55, .4, .35, .1])
p, r = pr_curve(y, s)
ap = np.sum((r[1:] - r[:-1]) * p[1:])          # average precision (step-wise)
plt.step(r, p, where="post"); plt.xlabel("Recall"); plt.ylabel("Precision"); plt.title(f"AP={ap:.2f}"); plt.show()

• Sort by score descending and use the cumulative TP/FP counts. AP is the area under the step curve. Mention sklearn.metrics.precision_recall_curve for checking.

• KNN: non-parametric; predicts from the k nearest neighbours.
• Simple, no training, handles multi-class.
• But: slow at prediction (O(n·d)), sensitive to scale and irrelevant features, suffers from the curse of dimensionality. Choose k by cross-validation.
• PCA: linear projection onto the directions of maximum variance.
• Reduces noise and dimensionality, speeds up models, aids visualisation.
• But: linear only, the components are hard to interpret, it is scale-sensitive, and it ignores labels.
• Together: PCA before KNN often helps, because distances become more meaningful in fewer dimensions.

• Context window: the maximum number of tokens the model attends to (prompt + output). Longer windows are costlier and can suffer "lost in the middle".
• KV cache: stores past attention keys and values during generation, so each new token needs only incremental compute. Memory grows with sequence length × layers × heads.
• Temperature: scales the logits. Lower is more deterministic; higher is more diverse.
• Hallucination: the model is trained to produce plausible continuations, not verified facts, and gaps in its knowledge or ambiguous prompts lead to confident errors. Mitigate with grounding (RAG), tools and verification.
• Fine-tuning vs RAG: fine-tuning teaches behaviour and format; RAG supplies fresh, citable knowledge.

Answer (STAR):

• Situation: "My teammate wanted to fine-tune a model; I believed RAG would meet the deadline and keep the data fresh."
• Action: "I suggested we each prototype for two days on the same 50-question evaluation set, and we agreed on the metrics in advance."
• Result: "RAG scored 0.86 against 0.81 for fine-tuning, at a fifth of the cost. We shipped RAG and used his fine-tuning work later for a classification step."
• Microsoft values a growth mindset and collaboration. Show that you listened, used data and kept the relationship strong.