Engineering work

7 projects shown

• 

  TwinTag ↗

  Co-built an industrial digital-twin workflow that turns scan imagery into reviewable 3D equipment tags.

  Technical details

  Problem: Equipment observations across industrial scans need to be localized, deduplicated, and linked to supporting evidence and documentation.

  Approach: Built for the VEO360 challenge at JunctionX Vaasa. A YOLO11m detector trained on synthetic data feeds E57-based 3D localization; pretrained Qwen3-VL extracts context for review in a Next.js workspace. The recorded demo uses precomputed backend results.

  Outcome: Demo: 108 scan images, 38 observations, 7 fused asset-tag proposals

  Python / YOLO11m / E57 / Qwen3-VL / Next.js / TypeScript

• 

  Qt Docking Contribution ↗

  Merged Qt Advanced Docking System fix preserving the globally focused dock widget when restoring a saved layout.

  Technical details

  Problem: Saved layouts retained each area's selected tab, but lost the globally focused dock widget after restoration.

  Approach: Persisted the focused widget's objectName in the docking state and reapplied focus after normal restoration completed.

  Outcome: Merged upstream on 19 August 2026

  C++ / Qt / Open Source

• 

  WardView ↗

  Responsive bedside-monitor interface concept built with Qt/QML and C++ models.

  Technical details

  Problem: Device-oriented interfaces must preserve state, remain readable across layouts, and communicate system conditions clearly without becoming a static visual mockup.

  Approach: A public Qt 6 project that separates QML presentation from C++ model ownership while exploring responsive layouts, reusable components, alarms, trends, and explicit interface states.

  Outcome: Public implementation with an intentionally bounded systems-learning scope

  C++17 / Qt 6 / QML / CMake

• 

  IPython Contribution ↗

  Merged upstream fix restoring pathlib.Path support in IPython IFrame rendering.

  Technical details

  Problem: An HTML-escaping security improvement passed pathlib.Path objects directly to html.escape(), causing IFrame rendering to fail and breaking downstream integrations.

  Approach: Reproduced the regression, traced the failing conversion boundary, converted the source to a string before escaping, and added focused regression coverage that also verifies special-character escaping.

  Outcome: Merged for IPython 9.16 after local and upstream validation

  Python / pytest / Git / Open Source

• 

  iMeal ↗

  Plan a week of meals from your recipes, turn the plan into a practical shopping list, and cook one step at a time.

  Demo ↗

  Technical details

  Problem: Recipes, weekly meal decisions, and grocery notes often live in different places, making it harder to plan and shop for the week.

  Approach: A private recipe library feeds a seven-day planner with four meal slots per day. Plans generate editable shopping lists, with AI-assisted ingredient organization reviewed before changes are saved. Shared lists and a focused step-by-step cooking view carry the workflow through to the kitchen.

  Outcome: Weekly planning, Shared shopping lists, Guided cooking

  Next.js / TypeScript / Tailwind CSS / HeroUI / Supabase / Zod

• 

  LLM Internals Lab ↗

  Learning lab implementing GPT-style building blocks from scratch, with public code for tokenization, vocabulary building, and next-token data sampling.

  Technical details

  Problem: It is easy to use AI APIs without understanding the layers underneath them. This public lab forces me to study implemented components directly rather than claim understanding from API usage alone.

  Approach: The current public tree contains a simple tokenizer, vocabulary builder, GPT dataset, and tests for tokenization and data sampling. Embeddings, attention, pretraining, and fine-tuning are the broader direction; they are not implemented in the current public tree.

  Outcome: Public tokenization and data-sampling implementations with tests

  Python / PyTorch / Tokenization / Data Sampling / pytest

• 

  Matchinary ↗

  Matchinary turns calendar screenshots into Google Calendar events with AI-powered OCR and calendar automation.

  Technical details

  Problem: Many calendars lack Google Calendar integration. Manual event entry from screenshots is slow and error-prone.

  Approach: Flask backend deployed on Render processes calendar images with AI-powered OCR. Next.js frontend on Vercel provides intuitive upload interface. Both communicate via REST API for seamless event extraction and Google Calendar sync.

  Outcome: Instant sync, AI-powered extraction

  Flask / Next.js / Python / AI/OCR / REST API / Google Calendar API / Render / Vercel

Professional portfolio →