An Interview with a FAANG engineer

Feedback about Mystic Thunderstorm (the interviewee)

Advance this person to the next round?

No

Strengths and what went well * Genuine enthusiasm for ML and hands-on building; clear ownership of an ASL→speech project. * Practical pivoting: moved from a CNN approach to a Mediapipe+MLP pipeline when overfitting/generalization became an issue. * Sensible feature engineering: derived ~160 features (distances, angles, finger crosses, palm-relative measures) from hand landmarks. * Thoughtful dataset hygiene: used cosine similarity to detect synthetic–real domain mismatch and discarded low-quality synthetic data. * Professionalism and growth mindset: honest about gaps, receptive to hints, and reflective about leadership experience and team dynamics. * Real-world exposure: evaluation/ranking work and regular use of dev tools (Copilot, etc.) show initiative. Areas of improvement * Core ML fundamentals: could not explain gradient descent/backprop, training loops, or bias–variance/overfitting mechanics beyond intuition. * Supervised/unsupervised breadth: limited recall of common unsupervised methods and when to use them. * Data/label representation: confusion describing a “row”/example and label encoding; need crisper articulation of inputs, targets, and shapes. * Model selection & justification: rationale for MLP architecture (layers/units/activations) and hyperparameters was ad-hoc; no mention of tuning strategy. * Evaluation rigor: no clear metrics (e.g., accuracy per class, confusion matrix), validation split/cross-validation, or error analysis beyond anecdotes. * Regularization & generalization: partial recall (dropout, early stopping) but not L1/L2, data augmentation strategy, or simpler-model baselines. * Communication under questioning: several clarifications needed; answers sometimes meandered before landing on the point. Advice for future interviews * Shore up fundamentals (target: crisp 1–2 minute explanations): * How training works: forward pass → loss → gradients via backprop → parameter update via (stochastic) gradient descent/Adam. * Losses (CE/MSE), activations, over/underfitting, regularization (dropout, L2), early stopping, data augmentation, bias–variance. * Common algorithms: k-means/GAUSSIAN MIXTURE/DBSCAN/PCA (unsupervised), logistic/linear regression, SVM, trees/ensembles, basic CNN/RNN/Transformer blocks. * Prepare a tight project narrative: * Problem, data schema (one row = 21 hand landmarks × (x,y[,z]) → engineered features; label = 26 classes, one-hot encoded), model architecture, training setup (optimizer, LR, epochs, batch size), metrics, results (overall and per-class), top failure modes, and next steps. * Add quantitative evidence: * Bring numbers: dataset sizes, train/val/test split, accuracy/F1, confusion matrix highlights (e.g., M vs N vs R confusions), ablation (with/without features). * Practice explaining label encoding and shapes: * Be ready to state: input tensor shape, output vector size, loss used, and why. * Strengthen generalization tactics: * Try transfer learning (e.g., MobileNet/ResNet) with augmentation (rotate/scale/translate, lighting), and compare against Mediapipe+MLP; use k-fold or a clean val set. * Build a repeatable workflow: * Simple hyperparameter tuning (grid/random), learning-curve plots, and training/validation metrics logged (TensorBoard). * Communication drills: * Mock Q&A on “row of data,” “why overfitting?”, and “how does backprop work?” Record yourself and tighten to concise, jargon-correct answers. * Have a leadership STAR story ready: * Situation/Task/Action/Result from this project—highlight driving consensus, timelines, and measurable outcomes. With these steps, you’ll present as both a doer and someone who understands the “why” behind the “what.”

Feedback about Admiral Hex (the interviewer)

Would you want to work with this person?

Yes