Smitkumar
Velani

A distributed ML training and inference platform spanning PySpark preprocessing, Ray-orchestrated PyTorch experiments, MLflow tracking, ONNX Runtime serving and Kubernetes deployment.

Processed 2.4M records and 42 features, cut a 24-run hyperparameter search from 96 to 29 minutes, and reduced p95 inference latency from 42 ms to 24 ms across 10,000 benchmark requests.

A five-service distributed execution platform combining Django REST APIs, PostgreSQL persistence, RabbitMQ/Celery asynchronous workers, Redis caching, Docker and AWS-oriented Terraform infrastructure.

Verified end-to-end asynchronous execution, runtime idempotency, persisted job state, explicit failure handling and 9/9 automated tests, with GitHub Actions CI.

Covers the full lifecycle of an experiment: design it, monitor it without inflating error rates, estimate the effect afterwards. CUPED, DiD, PSM, doubly robust AIPW, Bayesian A/B, uplift modelling and always-valid sequential inference, each validated against a known injected effect.

A placebo test measured a 0.45 per period pre-trend from pre-treatment data alone, then predicted the realised DiD bias of 0.37. Validated externally on the LaLonde benchmark, recovering 1,637 USD against a 1,794 USD experimental truth from a negative 8,498 USD naive baseline.

An evaluation harness built around Scout, a ReAct-style CSV analyst. Dual LLM judges score against a hand-written 29-task golden set, then get validated against human scoring, because a judge that scores answers is itself a model that can be wrong.

Red-teaming exposed a hallucination keyword guardrails could not see: the agent invented a customer segment absent from the data. I built a grounding guardrail validating entity mentions against the dataset itself, then documented the fabrication it still cannot catch.

Upload any CSV and it inspects data quality, flags target leakage, infers the problem type, trains five model families unattended, explains the winner with SHAP, checks for drift and writes a plain-English PDF report.

It selects the winning model by expected business cost rather than raw accuracy. On 284K rows of fraud data it chose XGBoost over Logistic Regression despite comparable AUC-ROC, because Logistic Regression's 0.06 precision was ruinously expensive. A concrete case for why metric choice matters on imbalanced data.