Aldrich Bañas

Personal Project

• Scoped and built an end-to-end computer vision tool that turns climbing ascent videos into automated, LLM-driven personalised coaching feedback, matched against a real board hold-layout database.
• Pose estimation and hold-contact detection with OpenCV and MediaPipe, reaching 83.7% accuracy through iterative testing with a custom calibration tool.
• Hybrid data model combining the boardlib/Aurora API with a manual-data fallback, keeping every feature available under network restrictions.

OxHack 26, University of Oxford — Tiger Data Track Winner

• Architected the medallion data layer for an AI that plays Pokémon battles in real time, cleaning 2.5M+ raw replays with DuckDB SQL into Parquet and loading them into time-partitioned TimescaleDB hypertables on Tiger Data.
• Shared the dataset across the team through read-only roles and optimised a key join from 133 s to 13 s.
• Engineered the offline datasets the agent loads to decide within a 3 s budget: probabilistic set guesses for 243 Pokémon from Smogon usage stats, a leakage-aware 490K-state training set and a validated 60-team test pool, cutting fallback decisions from 8.5% to 0.8%.

University of Dundee Hackathon — 1st Place

• Led the design of an agent-based market simulation, defining the requirements for dynamic pricing across 150+ autonomous agents driven by strategy and stress metrics.
• Real-time engine on an async FastAPI backend (asyncio tick loop) clearing supply and demand into live prices, with a Next.js / React / TypeScript frontend.
• Reusable component library (shadcn/ui, Radix UI, Tailwind CSS) and a typed React context store simulating unit swaps, payments and streaming loans with fee logic.
• Prioritised ruthlessly to pitch a complete MVP within the hackathon timeframe, winning 1st place against competing university teams.

Anthropic Claude AI Hackathon (Claude Builders Club) — People's Choice

• Identified the friction in crowdsourcing audio data and built a web portal where anyone can record their regional accent in the browser (Web Audio API & MediaRecorder), explored on an interactive Mapbox GL globe.
• Designed the metadata framework and a Node.js ETL pipeline covering 30+ languages from Wikimedia Commons and Hugging Face Datasets, backed by Supabase.
• Won People's Choice by presenting a clear product narrative and a live interactive demo focused on cultural preservation.

Hardware Product Design Project

• Designed and built a handheld reaction-time game for patients with degenerative brain conditions, taking it from user needs through CAD, prototyping and user testing.
• Modelled the enclosure in SolidWorks and fabricated it with CNC machinery; designed and carved a custom PCB, then wired and soldered the LED targets, two-digit display and buzzer.
• Programmed the game logic as flowcharts in Circuit Wizard.
• 100% positive feedback from user testing, with 92% of testers requesting no changes.