Aditya Dogiparthi

OpsLens — Verifiable AI for Operational Investigation

Built an assistant over synthetic production records and authored runbooks that identifies output losses, retrieves relevant procedures, and returns executed SQL, charts, and supporting evidence. • Achieved 92% numeric-answer accuracy on 200 held-out, answerable questions scored against reference SQL; separately tested handling of missing evidence, duplicate joins, inconsistent units, and delayed events. • Reduced median time to complete operational analysis tasks from 18 to 7 minutes in a counterbalanced prototype study with 12 student participants, compared with a dashboard and manual-query workflow. • Implemented read-only queries, validation checks, workflow checkpoints, and execution traces; documented incorrect answers and measured response latency and inference cost.

F1 Race Winner Prediction — ML Pipeline for Race Outcome Forecasting

Machine learning pipeline that predicts Formula 1 race winners from historical race results, driver statistics and weather conditions, using automatically scraped race data. • Built web-scraping and Pandas data pipeline collecting historical race results, driver statistics and weather conditions into a cleaned, model-ready dataset for training and live prediction. • Engineered features and trained Scikit-learn and XGBoost classifiers to predict race winners, comparing models with a heldout train-test evaluation to select the best performer. • Evaluated models on unseen races, reaching 76% winner-prediction and 89% podium-prediction accuracy on 2024 season race

LineSight — Industrial Defect Detection with Limited Training Data

• Built anomaly detection and localization from frozen DINOv3 features, nearest-neighbor scoring, and 50 defectfree reference images per category; evaluated performance as reference data became more limited. • Lowered the false-positive rate from 12% to 8% at 90% defect recall against a PatchCore baseline across four MVTec AD 2 category evaluations. • Achieved 180 ms p95 inference latency for 256 × 256 images on an Intel Core i7-11800H CPU; documented how image resolution and model optimization affect small-defect detection. • Developed an inspection interface showing anomaly heatmaps and comparable normal images; reported categorylevel failures under lighting changes that require human review.