CONNOR LIN

The architecture is a deterministic OCR spine with routed VLM overlays — page-level quality scoring identifies problematic content and routes only those pages to heavier models, so expensive stages run only where they provably win. Computer vision stages recover checkbox states from degraded scans (faded borders, hand-drawn ticks, bleed-through) and de-weld superscript footnote markers that OCR engines misread as digits in financial figures — a class of silent corruption where a single welded character can turn a correct dollar amount into a plausible 10× error. An independent OCR cross-check catches silent line-detection misses from the primary engine. Engine selection was settled through structure-aware bake-offs across 5+ OCR/VLM candidates, and a CI-gating evaluation harness enforces exact match on financial fields and label–value association against hand-verified ground truth. The pipeline runs deterministically and local-by-default on consumer GPUs, with serverless GPU offload for heavy stages and reproducible parity across dev and production environments. (Proprietary codebase — no public repo.)

The project employs a convolutional neural network (CNN) to classify STM tip sharpness, with custom image processing and contour extraction. Data augmentation techniques bolster model resilience despite a limited dataset. An ROI-based cross-scan algorithm further enhances accuracy, ensuring reliable classifications and forming the foundation for full automation in STM-based fabrication.

The app is built on Tauri 2 with a Rust backend running on Tokio and a React/TypeScript frontend. The backend watches Valorant's lockfile for the local client port and credentials, holds a persistent websocket for game-state events, and decodes the player's presence to track the Menus, Pregame, and Ingame states. On a state change it fetches the match roster, then fans out per-player calls for MMR, competitive history, match details, and loadouts. That enrichment is rate-limit safe by design: a fixed 120 ms inter-request delay, Retry-After backoff on 429s, and session-lived caches keyed by match ID and game phase, which brought a full 10-player lobby from about 15 seconds with frequent 429s down to about 5 seconds with none. A fast snapshot (names, ranks, agents) reaches the UI in under a second while heavier stats fill in asynchronously, and a cancellation signal checked between every API call aborts stale work when the player leaves agent select. The backend owns row ordering and privacy enforcement, and publishes a new snapshot only when content actually changed. Releases are built by GitHub Actions on tag push, and 146 unit tests cover the parsing and aggregation logic.

The system ships as two compiled Go binaries (host and guest) with role baked in at build time, plus an mpv Lua script and config file. The helper exposes a local HTTP API that the Lua client polls asynchronously with bounded timeouts so playback is never blocked by network issues. Room state lives in Firebase Realtime Database — the host writes playback position, pause state, and server-calibrated timestamps, while guests read and apply corrections. Adaptive polling adjusts the interval between idle and active states. Seek-lock prevents guest drift beyond a configurable gap, and auto force-sync triggers when the host makes large seeks. The host dashboard is a vanilla HTML/JS/CSS app served by the helper at localhost:8765, with shared room-event toasts and per-room settings (poll intervals, seek lock gap, cooldowns) that persist in Firebase. Cross-platform builds target Windows (host + guest) and macOS Apple Silicon (guest) via PowerShell and shell build scripts, with .env-driven Firebase URL and room defaults baked into release bundles.

The backend is entirely developed in Python and utilizes the A* pathfinding algorithm to calculate the shortest path between two points. It works by converting a picture of the school map into a 2D array, with each pixel representing a node. The program then applies the A* algorithm to find the shortest path between two nodes and incorporates a breadth-first search algorithm to determine the necessary locations to traverse in reaching the destination.

Surplus Share utilizes a Python backend with a heuristic genetic algorithm to optimize delivery routes for drivers. This algorithm leverages the Google Maps Distance Matrix API to calculate travel times and distances between locations, taking into account real-time traffic conditions. Additionally, a distribution algorithm ensures efficient allocation of food resources based on dietary restrictions. Firebase Firestore serves as the database, storing information about businesses, individuals, and available food. The frontend, built with Nuxt.js and TailwindCSS, provides a user-friendly interface for both businesses and individuals.