Building systems that
ship and scale.

Built a full-stack interview simulator that uses dual LLM agents — one for real-time answer evaluation against weighted rubrics, one for neutral interviewer delivery — to conduct adaptive Data Science interviews over WebSocket. The engine groups questions by domain (depth-first), branches follow-ups based on answer quality, enforces realistic pacing (~4 min/question), and generates post-interview reports with 6-dimension scoring and personalized 7-day training plans.

Designed a dual-agent LLM architecture that separates evaluation from delivery, enabling realistic interviewer behavior — neutral probing, adaptive difficulty, and time-pressure cues — across 100+ questions in 10 DS domains.

Implemented and evaluated structure-aware medical image segmentation workflows for retinal OCT analysis, focusing on improving segmentation quality under challenging clinical imaging conditions.

Improved segmentation performance and strengthened experience translating foundation-style vision models into medical AI settings.

Investigated the limits of CNN-based detection and tracking under real-world camera conditions — then independently designed a transition from bounding-box tracking to pixel-level segmentation. Started with YOLO + DeepSORT + Kalman filtering for fixed-camera scenarios, identified systematic failure modes in moving-camera settings (ID switching after occlusion, lost small/edge objects), and drove the architectural pivot to transformer-based segmentation (SegFormer, Mask2Former, SAM) to recover scene understanding where tracking broke down.

Improved mAP from 19% to 43% by diagnosing tracking failures and benchmarking segmentation architectures as a principled alternative.

Designed and analyzed an A/B test comparing a control (current free trial flow) against a treatment (time-commitment screener) to reduce cancellations and improve retention. Evaluated click-through, gross conversion, and net conversion using t-tests and confidence intervals, performed sanity checks on invariant metrics, and recommended launching based on statistically and practically significant retention improvements.

Validated a retention-improving intervention with statistical rigor, recommending launch with follow-up tests to address early cancellations.

Developed interpretable ML models (Logistic Regression, Random Forest, XGBoost) on 10K loan records with SMOTE and threshold tuning for class imbalance. Applied SHAP, LIME, and permutation feature-importance to surface top default drivers and deliver regulator-ready explanations. Audited fairness across borrower demographics to inform equitable loan-approval policies.

Achieved 92% recall on defaults with PR-AUC of 0.84, while maintaining demographic parity in false-negative rates.