LLM agents, retrieval systems, forecasting, and ML-accelerated simulation.
Selected machine learning work spanning production LLM agents, retrieval systems built from scratch, applied forecasting, and ML surrogates that accelerate first-principles numerical solvers.
Deployed an adaptive LLM agent for Python education: generates a personalised learning roadmap per student, tracks progress in real time, and produces structured reports for students, parents and teachers — end-to-end from interaction data to deliverable output. Built the learning-map engine sequencing lessons to each student's level, iterated across diverse learner profiles.
Built a retrieval-augmented QA engine combining BM25 lexical scoring with dense embedding similarity, fused with reciprocal rank fusion, exposed through a Streamlit query interface — every scoring component hand-written rather than assembled from a framework.
View repository →Full pipeline from source to generated text: automated scraping via Playwright (Cointelegraph, Decrypt, CoinDesk), memory-efficient preprocessing with HuggingFace Datasets, and domain-specific language-model fine-tuning for automated commentary.
A staged LLM pipeline that enriches and validates HS (Harmonized System) customs codes at scale: prompt generation, LLM inference, rule-based post-processing and a Streamlit review step, run as independently testable phases rather than one script.
Engineered a forecasting suite (ARIMA, Prophet, Exponential Smoothing, XGBoost, LSTM, Gaussian Processes) for bike-share demand — a 12% accuracy improvement enabling a 25% cost reduction in fleet redistribution.
Random Forest, XGBoost and ANN churn models achieving a 15% improvement in retention-strategy effectiveness, deployed as scoring APIs for downstream product use.
Built causal-inference models on large-scale behavioural data to estimate true effect beyond correlational readouts, plus anomaly-detection systems flagging irregular user and traffic patterns; designed A/B tests and automated analytics dashboards for product and engineering stakeholders.
CUDA C++ (NVCC / C++17) engine from first principles: cubic equations of state (SRK/PR), multiphase flash, stability analysis, phase-envelope construction, excess-Gibbs mixing rules (Huron-Vidal/NRTL). Now developing ML surrogate models of the flash solver to cut iterative cost toward real-time flow simulation — validated to production tolerances against a commercial reference; export pipelines feed LedaFlow and PIPESIM for production, flow-assurance and CCUS applications.
Regression + dimensionality reduction over DNS-LES boundary-layer data, cutting simulation time from several weeks to a single day. Demonstrated a potential 10% drag-force reduction with direct implications for aerospace fuel cost, contributing to peer-reviewed output.
Source behind the projects above — LLM pipelines, retrieval, forecasting, and applied ML experiments — plus course work and notebooks.
View repository →More background on skills & tools · full history on LinkedIn · code on GitHub