I'm Ekansh Mittal

Engineered a Neural Network with Bayesian Optimization to predict breast cancer drug response, achieving 92% accuracy on clinical trial data. Implemented SHAP and ExtraTrees for feature selection, identifying IGF1R as a key resistance marker from >16,000 gene expression features. Validated predictions by developing a 3D spheroid co-culture model, proving a computationally identified triple-therapy regimen reduced cell growth by 3-fold.

Built a multimodal PPO trading agent that makes decisions by processing live earnings call transcripts synchronized with high-frequency stock data. Evaluated BERT and Longformer architectures, identifying a trade-off where long-context models provided higher upside (>4.6% max return) but lower stability than static baselines.

Developed SkyBites, a PPO-based RL system to coordinate 25 autonomous drones, outperforming greedy baselines with a 50% reduction in delivery delay via long-term horizon planning. Implemented Action Masking and Curriculum Learning to handle complex constraints (battery, payload) and scale training to 1,000 stochastic daily orders.

Built LSTM and Feed-Forward Neural Networks to predict metabolic rates from sensor data, achieving 0.0034 MSE on 30-minute forecasts. Processed 45 days of high-frequency biometric data (CGM & Calorimetry), utilizing SHAP to identify body temperature as a primary driver of glucose variability.

Developed a novel microfluidic Device to coculture bacteria and cancer cells. Identified microbial signatures and key drivers of regulation of cancer cell growth. Patent Pending.