All case studies

Solves the "one-size-fits-all" education problem by adjusting question difficulty and learning paths based on real-time performance, mimicking a private tutor.

• Built an adaptive learning map (Duolingo-style) where subtopics unlock sequentially as students prove mastery through real-time BKT assessment.
• Designed a Teacher Command Center with class-wide heatmaps and automated struggle alerts that flag at-risk students for timely intervention.
• Engineered a custom Bayesian Knowledge Tracing engine with logistic extensions, time decay, consistency gating, and adaptive difficulty selection.

Lets visitors explore the portfolio conversationally and demonstrates end-to-end RAG, prompt constraints, deployment, and admin operations.

• RAG over knowledge files + Django DB content
• Admin-controlled maintenance mode, tracking, and index refresh
• Reduced redundant LLM calls for standalone questions
• Optional FAISS persistence for faster restarts
• Excel Q&A logging for evaluation/testing

Led research and engineering for an AI system assisting hemodialysis teams with anemia management, pairing OCR pipelines with role-aware dashboards.

• Architected a robust Vision AI pipeline using Google Gemini to extract context-aware structured data from medical tables with 85-95% accuracy.
• Engineered a high-resilience batch engine with a custom API Key Rotation & Failover system to handle high-volume processing and rate-limit bypass.
• Integrated a longitudinal analytics dashboard that tracks 20+ biomarkers over time, helping clinicians identify health patterns 90% faster.

An AI-powered clinical decision support system designed to automate anemia management for hemodialysis patients.

• Directed a multi-member team, owning experimentation, model selection, and delivery.
• Combined OCR and ML inference to extract lab metrics with reduced manual errors.
• Built secure experiences for doctors, nurses, and patients plus a Flutter companion app.

Designed custom NLP pipelines for classification, similarity, and generation.

• Designed custom NLP pipelines for classification, similarity, and generation.
• Trained FastText models (custom + pretrained), improving Yelp dataset word similarity tasks.
• Built LSTM & GRU-based models for text generation on Wikipedia data.

Implemented Q-Learning & Value Iteration on OpenAI Gym’s FrozenLake-v1.

• Implemented Q-Learning & Value Iteration on OpenAI Gym’s FrozenLake-v1.
• Developed optimal policies for both stochastic and deterministic versions of the environment.
• Compared performance of algorithms in stochastic vs deterministic environments.

Developed a VAE with reparameterization trick for robust latent space sampling.

• Developed a VAE with reparameterization trick for robust latent space sampling.
• Trained on CIFAR-10 RGB and a custom dialysis dataset linked to graduation project.
• Created image interpolation functions for smooth transitions.
• Extended model into a GUI tool for interactive image generation.

Evolved an autonomous Tetris agent that surpassed 10.3M points via GA-driven heuristics.

• Implemented selection, crossover, and mutation strategies to iterate on candidate agents.
• Instrumented training loops and telemetry to surface board-evaluation insights.

Consolidated fragmented fitness workflows into a single web app, pairing real-time training analytics with scientific nutrition guardrails and coach-client collaboration.

• Built an automated ingestion pipeline to maintain a 1,700+ exercise library with media and structured muscle mapping.
• Implemented a set-level logging system with real-time Personal Record detection and dashboard visualization.
• Delivered a scientific nutrition engine (Mifflin–St Jeor) with adaptive daily macro guardrails and guest-mode conversion.
• Enabled coach-client workflows for assigning and updating training plans remotely.