Nimesh Kulatunga

CS3203 Software Engineering Project at University of Moratuwa, built with a team of 3. Own the data layer: an ETL pipeline over CSE data that updates daily in CI, with validation for OHLC consistency, duplicates, and coverage gaps, backed by a test suite and quality reports.

Built an evaluation harness and generation pipeline comparing a baseline against shape-aware prompting for CUDA kernel generation.

Completed the Summer of Bitcoin 2026 Developer Track challenge round by implementing Chain Lens, a Bitcoin transaction and block parser; Coin Smith, a safe BIP-174 PSBT transaction builder; and Sherlock, a Bitcoin chain-analysis engine with heuristic reports and a web visualizer.

Includes a settlement engine with ACID-compliant, double-entry bookkeeping, and a transaction router built for sub-millisecond routing with circuit breakers and netting.

Built with a team of 5. The MVP uses a Raspberry Pi that polls the vehicle's ECU every second and posts batched telemetry to a FastAPI backend, feeding a dashboard that computes fuel and driver-efficiency KPIs from the logged data.

Built a sophisticated trading agent using Deep Reinforcement Learning that learns optimal portfolio allocation strategies. The agent manages a diversified portfolio of 10 assets using PPO algorithm, technical indicators, and market stress features. After 10 systematic experiments, achieved a Sharpe ratio of 1.617 and deployed to Alpaca paper trading.

Complete NLP pipeline benchmarking BERT, RoBERTa, DistilBERT, and FinBERT on financial sentiment classification. FinBERT achieved 87.2% accuracy, outperforming general-purpose models by 3-6%. Extended to news-driven trading signals with proof-of-concept backtesting. Includes live Streamlit dashboard on Hugging Face Spaces.

Built a "Council" of specialized LLMs (Qwen2-Math, DeepSeek-Coder, Llava) that runs entirely offline on an 8GB RAM laptop. Features a custom weighted-keyword router (96% accuracy), RAG with 2,600+ textbook pages, and real-time DuckDuckGo web search. Accessible via a Gradio web interface.

Comprehensive implementation of Stanford Pintos operating system for CS2042 coursework. Implemented advanced interactive shell with 15 commands (performance monitoring, heap analysis, scheduler inspection) and complete user program support with 13 system calls, process management, and file system operations using semaphores for synchronization.

Developed a comprehensive web platform that connects elders and their families with verified and qualified caregivers. Worked on backend development and database design to ensure a robust and scalable solution.

Developed an 8-bit nanoprocessor as part of computer organization and digital design coursework. The project involved implementing instruction set architecture and control unit design, demonstrating understanding of fundamental computer architecture principles.