Architected a microservices-based agentic AI research pipeline with a PostgreSQL + Redis RAG system, achieving 40% faster query response times. Integrated Google Cloud Vision API for multimodal document analysis supporting 10+ file formats with 85% extraction accuracy. Deployed Dockerized services with CI/CD pipelines, achieving 99.9% uptime. Wrote Jest unit tests covering RAG retrieval, database seeding, and connection validation.

Fine-tuned DeepSeek-R1-Distill-Llama-8B on financial datasets using LoRA with Unsloth, achieving a 15% performance gain over the base model on domain-specific financial Q&A benchmarks. Built an end-to-end training and evaluation pipeline. Model reached 4,100+ downloads on Hugging Face and 15+ GitHub stars.

Built a REST API in Spring Boot to shorten and expand URLs, storing mappings in PostgreSQL with Redis caching to reduce repeated database lookups. Implemented per-IP rate limiting using Redis and async hit count tracking with a scheduled aggregator to minimize write load on PostgreSQL.

Built an ETL pipeline ingesting 1,000+ daily sales transactions with data quality checks into a PostgreSQL star schema warehouse, reducing query latency by 60%. Wrote 15+ analytical SQL queries using window functions and CTEs. Trained 3 forecasting models (Linear Regression, Random Forest, XGBoost); best model achieved 36% RMSE improvement and 91% R² score. Connected Power BI dashboards to PostgreSQL for stakeholder reporting.

Implemented MRI-to-CT cross-modality translation using CycleGAN with ResUNet generators, achieving a 0.65 SSIM score on held-out clinical data. Designed a custom composite loss combining adversarial, cycle-consistency, and SSIM terms. Built a MONAI-based preprocessing and augmentation pipeline for clinical evaluation.

Built an HTTPS server in C using OpenSSL and BSD sockets, performing TLS handshakes and handling HTTP requests over secure TCP connections on Linux. Implemented signal handling and deterministic cleanup of sockets and SSL resources to ensure stable shutdown and memory safety.

Developed a Gemini-powered mock interview tool providing real-time feedback to enhance user preparation. Implemented posture tracking with TensorFlow.js to alert users when off-frame. Engineered a backend with Supabase and Clerk, securing user data and cutting authentication times by 40%.

A web application that allows users to upload legal documents (PDFs or images) and perform intelligent, context-aware queries on those documents.

Transforms website content into vector embeddings, summarizes it, and generates customized advertisement images based on user prompts. Built during Rocketium's #CreatiVerse 30hr Hackathon — top 8/50+ projects.