Saransh Surana

Data Science - ML - AI

Saransh • Data Science • AI

I ship AI systems that actually get used.

From messy data to deployed models – I care less about leaderboard scores and more about shipped systems that move metrics for real teams.

Python • LLMs & RAG • MLOps • Experimentation • Evaluation

Ford (Data Science), Stony Brook (MS Data Science), and a handful of startups and research teams where I shipped models instead of just slide decks.

Built RAG-based structured extraction with JSON Schema guardrails, lifting parse success from 78% to 96% and cutting hallucinations by 25%.
Ran 50+ MLflow experiments across models, prompts, chunking, and retrieval to systematically tune extraction quality.
Shipped a FastAPI + Celery + Redis + PostgreSQL pipeline that processes 1.2–2k records/hour with under 2% timeouts and 35% fewer repeat LLM calls, plus a React + TypeScript directory UI with search and filters.

PythonRAGGemini FlashMLflowFastAPICeleryRedisPostgreSQLReact.jsTypeScript

Built Python pipelines to clean and preprocess unstructured data from web pages, PDFs, and other raw formats, version-controlled with Git for reproducibility and collaboration

PySparkNumPyPandas

Built scalable ETL pipelines in BigQuery and GCP to support end-to-end ML workflows for anomaly detection on manufacturing sensor streams.
Trained Isolation Forest and One-Class SVM models that achieved 78% recall and 73% precision for early fault detection.
Turned anomaly insights into fixes by explaining model behavior and findings to both engineering and leadership teams.

Pythonscikit-learnPandasmatplotlibGCPAnomaly Detection

Deployed XGBoost demand-forecasting models on GCP with FastAPI and Docker, reducing forecast error by 18% across 30+ SKUs.
Designed Spark / Hive / Kafka pipelines and SQL-based ETL that cut deployment time by 40% and supported 1K+ events per day.
Lifted marketing ROI by 30% through clustering 10K+ customers into high-value segments and validating promotions with A/B tests (+7% conversion).

XGBoostFastAPIDockerBigQueryGCPA/B TestingKafkaSparkHivescikit-learnClusteringMarketing