Mphasis & Persistent: GenAI / AI-ML Engineer

• Uses: GANs for super-resolution and style transfer; VAEs for anomaly detection and representation learning; diffusion for text-to-image, image editing and audio.

• Images:
• FID: the distance between real and generated feature distributions. Lower is better.
• Inception Score, precision/recall for generative models, and CLIP score for text-image alignment.
• Text: task metrics (exact match, ROUGE), LLM-as-judge, human preference, and faithfulness for RAG.
• Always include human evaluation on a sample, plus safety checks (toxicity, bias) and diversity versus quality trade-offs.

1. Check the data first: a feature pipeline mismatch between training and serving (training-serving skew), missing features, or schema changes.
2. Drift: compare production input distributions with training (PSI, KS test). Concept drift means the label relationship has changed.
3. Evaluation flaws: leakage in the offline set, or a random split when it should have been time-based.
4. Fix: retrain on recent data, add monitoring and alerts, run a shadow model, and roll back if needed.
5. Prevent: a feature store, data contracts, and a scheduled re-evaluation.

• Inputs: a user profile and segment, recent behaviour, product catalogue and brand guidelines.
• Pipeline: 1. A candidate generator: rules plus a recommender picks products and offers. 2. An LLM writes copy from a template prompt (tone, length, compliance rules), with structured output. 3. Guardrails: banned claims, PII, brand checks. 4. Human approval for new templates.
• Scale: batch generation overnight and caching by segment; real-time generation only for high-value triggers.
• Measure: an A/B test on open rate, click-through and conversion, plus unsubscribe rate as a guardrail.

• Event stream (Kafka or Kinesis) → stream processor (Flink or Spark Streaming) that computes features → online feature store (Redis) → model service (low-latency REST or gRPC, autoscaled) → decision / action.
• Keep offline training consistent with the same feature definitions.
• Latency budget: under 100 ms, using lightweight models, cached features and async logging.
• Reliability: fallback rules if the model times out, canary deploys, and monitoring of latency and drift.

import numpy as np

def synth(n, dist="normal", seed=42, **p):
    rng = np.random.default_rng(seed)
    if dist == "normal":  return rng.normal(p.get("mu", 0), p.get("sigma", 1), n)
    if dist == "poisson": return rng.poisson(p.get("lam", 3), n)
    if dist == "categorical":
        cats, probs = zip(*p["probs"].items()); return rng.choice(cats, size=n, p=probs)
    raise ValueError(dist)

print(synth(5, "categorical", probs={"A": .5, "B": .3, "C": .2}))

• For realistic tabular synthetic data, mention SDV (CTGAN or Gaussian copula), keeping correlations, and privacy checks.