What I've Built

Challenge

Financial data is messy. Multiple APIs return different formats, different timestamps, different precision. Existing charting libraries couldn't render real-time candlestick data at the speed and granularity required — they'd choke on the update frequency or produce visual artifacts during rapid redraws. On top of that, congressional trading data comes from scattered government filings that need to be parsed, normalized, and cross-referenced with live market data to be useful.

Solution

Built a custom charting engine from scratch using Canvas API for sub-second candlestick rendering with zero visual artifacts. Designed a normalization pipeline that ingests disparate financial data sources and unifies them into a single format with consistent timestamps and precision. Congressional trading and SEC filing data flows through a dedicated parser that cross-references holdings with real-time price data automatically.

Challenge

Generative AI models are resource-intensive and slow. Each modality — image, music, audio — has different processing requirements, different latency profiles, and different output formats. Existing solutions dump users into raw model interfaces with no iteration workflow, no output management, and no way to manage GPU costs at scale. The UX challenge was making three fundamentally different ML pipelines feel like one cohesive product.

Solution

Built an abstraction layer that normalizes the generation workflow across all three modalities — same input patterns, same iteration flow, same output management. Implemented a job queue system with real-time progress tracking and preview generation so users get immediate feedback during the 10-60 second generation window. Resource allocation dynamically routes workloads to available GPU capacity to minimize cost and maximize throughput.