Problem

Daily macOS workflows often require switching between full apps for music context, clipboard history, and quick notes, while the physical notch/menu-bar area remains underused as a persistent glanceable surface.

Approach

• Built a native macOS utility in Swift, SwiftUI, and AppKit with a floating top-center NSPanel that attaches visually to the physical notch or behaves as a virtual island on notchless displays.
• Separated testable app logic into TopNotchCore while keeping window lifecycle, status item behavior, and SwiftUI panel rendering in the app target.
• Implemented provider boundaries for Apple Music now-playing state, playback controls, lyrics fallback states, and future media-provider expansion.
• Added privacy-first local clipboard history with text-only capture, pre-persistence sensitive-content filtering, retention limits, and copy-back behavior that avoids duplicate churn.
• Created local Markdown scratchpad and pinned-note stores, display calibration settings, multi-display positioning, and planned module tiles for future calendar, timer, file-drop, commands, and read-only agents modules.

Outcome

A local-first macOS workflow surface that keeps music, lyrics, clipboard, notes, and future module context one click from the notch without accounts, cloud sync, telemetry, or background server storage.

Problem

Tailoring CVs and cover letters for individual job descriptions is highly repetitive, slow (often taking 2+ hours per role manually), and prone to copy-paste errors or misalignment.

Approach

• Developed a private AI Career Manager utilizing Python/FastAPI backend services and Pydantic contracts for structured data validation.
• Constructed a suggestions-first tailoring workflow that extracts structural metadata and requirements from incoming job descriptions.
• Designed grounding mechanisms with provenance metadata, unsupported-claim checks, and ATS/QA reporting to prevent LLM hallucinations.
• Created a React + TypeScript + Vite web interface where the user can review and edit suggested bullet points side-by-side with the original CV.
• Implemented template-aware document rendering using python-docx and pdfplumber to export clean files matching specific page budgets.
• Separated the agent orchestration layer (Eve) from deterministic Core services, keeping human-in-the-loop approval gates explicit.

Outcome

Reduces a 2+ hour manual resume tailoring process to minutes, ensuring strong alignment with job descriptions while guaranteeing factual accuracy and preserving final user control.

Problem

Querying internal corporate policy, public-policy, and procedure documents (often hundreds of pages across PDF/DOCX) is slow for operational teams and prone to misinterpretation or hallucination if standard LLMs are used.

Approach

• Engineered a Retrieval-Augmented Generation (RAG) pipeline with a FastAPI backend and a React/TypeScript/Vite frontend using Entra sign-in.
• Built an ingestion pipeline featuring SHA-256 file deduplication, PDF/DOC/TXT parsing, and Azure Document Intelligence OCR fallback for low-confidence or scanned text.
• Implemented top-k document retrieval using local embeddings and a FAISS vector database.
• Set a strict minimum-similarity score gate to decline answering when no relevant source document is found, enforcing "cite-or-decline" behavior.
• Integrated local Ollama models via LangChain components to process queries securely and locally without external cloud dependencies.

Outcome

A secure, pilot-ready ops chatbot that delivers reliable answers with exact file-path and page-number citations, reducing information discovery time while eliminating hallucination risk.

Problem

Digitizing, organizing, and rendering a large collection of family archive documents and high-resolution images in a responsive, fast-loading, and secure public platform without complex database maintenance.

Approach

• Developed a high-performance single-page web application (SPA) using Vite and Vanilla JavaScript/CSS.
• Created an interactive family tree visualizer with dynamic tree mapping, multi-criteria search, branch filtering, and a details drawer.
• Designed a responsive image delivery pipeline utilizing WebP/JPEG assets, gallery-specific loading states, and high fetch-priority for key visual elements.
• Implemented Firebase and Cloud Firestore for light data storage, visitor board interaction, and message collection.
• Added security measures including HTML escaping and domain allowlisting.
• Set up a deployment pipeline on Vercel with integrated analytics and release validation scripts verified via Puppeteer user-flow tests.

Outcome

A fast, accessible, and secure genealogy portal that displays rich historical data and family relations, maintaining responsive load times and high visual quality across devices.

Problem

Meal planning, recipe management, and grocery list generation are traditionally fragmented and manual tasks, leading to inefficient shopping, duplicate ingredient entries, and a high rate of household food waste.

Approach

• Developed a full-stack meal planning application utilizing React 19, TypeScript, Vite, Tailwind CSS 4, and Framer Motion for a responsive interface.
• Built an Express/Node.js backend with SQLite persistence via better-sqlite3 for fast local storage.
• Integrated Google Gemini and local Ollama models for automated recipe parsing, nutrition macro estimations, and aisle categorization.
• Implemented an intelligent grocery deduction engine that calculates required items from planned meals and subtracts current pantry inventory with unit-aware conversions.
• Created ingredient hygiene tools including a bulk merge utility and strict unit validation to consolidate duplicate ingredient entries.
• Designed clean PDF exports for generated grocery shopping lists.

Outcome

A unified, AI-enhanced kitchen companion that simplifies meal schedules and inventory tracking, eliminating database clutter and generating ready-to-use shopping lists.