Selected work, kept concise and safely summarized where the work is private.
A system for reducing manual reconciliation work in accounting teams. It combines spreadsheet parsing, deterministic matching, a review interface, and LLM-assisted review to separate clear matches, ambiguous cases, and client-facing questions. The business goal: faster, more auditable review without hiding the evidence.
A complex agent, retrieval, and document-processing system for an education business with a large body of specialized material. Built across Next.js, Node APIs, Python processing jobs, vector search, embeddings, reranking, and streaming LLM responses. The goal was not just a chat interface, but a maintainable way to turn internal knowledge into a searchable learning experience.
Production application for coordinating files, documents, statuses, human review, and handoffs in a document-heavy legal workflow. Built around React, FastAPI, async workers, storage, queues, and LLM-assisted extraction/review. The value was turning an artisanal process into a traceable, measurable, much faster operation without exposing domain details.
Product in development for operational teams moving from scattered spreadsheets to a clearer, reviewable, exportable working interface. Built with React/TypeScript, map-based UI components, spreadsheet import/export, PDF reporting, and careful state management. The emphasis is daily usability, constraint control, and reducing manual coordination rather than exposing domain-specific mechanics.
Production pipeline for receiving operational documents, extracting structured data with GenAI, validating results, and sending them into SAP. The important part was not simple classification, but turning messy documents into usable business records with fewer manual entries, fewer errors, and less friction between support, operations, and back office.
A set of fast-moving systems across business analytics, voice workflows, data products, and generative media. The common thread is translating ambiguous needs into working demos or handoff-ready systems with enough architecture, deployment, and documentation to be useful after the first presentation.
Open-source tooling around AI coding assistants and document workflows, including practical improvements for Markdown-based editing and larger document insertion. Small, utilitarian projects, but aimed at removing friction in real agent-assisted workflows.
Design of vaults and context systems that turn notes, meetings, projects, and decisions into useful knowledge bases for humans and AI agents. The work is less about organizing Markdown and more about structure, conventions, graph links, retrieval, and operational memory.
Master-level coursework around computational linear algebra: matrix methods, decompositions, numerical stability, and applied modelling. Especially relevant as mathematical infrastructure for optimization, machine learning, and data systems.
Master-level coursework on numerical and computational techniques for differential equations. It sits at the intersection of applied math, simulation, numerical approximation, and scientific computing.
Robotics projects with C++ and algorithmic work around perception/control-style problems. Relevant because it shows comfort with lower-level engineering and systems that interact with physical constraints.
Course projects around image processing and computer vision methods. Useful foundation for document/image extraction work and multimodal AI systems.
University work on mathematical optimization: modelling objective functions, constraints, numerical methods, and decision problems. Useful background for later operational planning and allocation systems.
Course projects around probabilistic modelling, uncertainty, priors/posteriors, and statistical inference. The relevant signal is comfort reasoning under uncertainty rather than treating model outputs as absolute facts.
Applied modelling work covering supervised learning, evaluation, model selection, and statistical trade-offs. A bridge between mathematical foundations and practical machine learning.
Projects implementing and experimenting with neural-network models, training loops, evaluation, and the practical details that make ML systems work beyond the formula sheet.
A small custom Gymnasium environment around Snake, useful as a practical reinforcement-learning exercise: environment design, reward shaping, game loop, and agent experimentation.