case studies
Curated Projects
MoodMetrics is a real-time cryptocurrency sentiment analysis engine that aggregates market sentiment from Reddit, CoinGecko, crypto news RSS, Binance funding rates, and the Fear & Greed Index. It uses AI (Nvidia NIM — Llama 3.1 8B) to analyze social posts and generate a blended sentiment score for major cryptocurrencies (BTC, ETH, SOL). Built with Next.js 16 and Drizzle ORM on Neon PostgreSQL, it features a tactical dashboard, AI Records viewer, strategy backtester, advanced quantitative tools (lag correlation, narrative clustering, whale vs. retail analysis), and a retro terminal-style loading screen — all wrapped in a premium dark UI with spring micro-interactions.
★Built a real-time multi-layer sentiment aggregation engine that fuses data from 5 independent sources — Reddit AI analysis, CoinGecko news, crypto RSS feeds, Binance perpetual funding rates, and the Fear & Greed Index — into a single blended sentiment score with configurable weightings (45/35/15/5).
★Integrated Llama 3.1 8B via Nvidia NIM for AI-powered sentiment classification, complete with a custom JSON fallback parser that extracts label, score, confidence, and reasoning from LLM output. Added pgvector embeddings for semantic similarity search across historical AI verdicts.
★Engineered a robust Reddit scraper with 8 anti-blocking defenses including exponential backoff, user-agent rotation, sequential execution with randomized jitter, Retry-After header respect, and an Arctic Shift fallback API — achieving reliable subreddit ingestion under aggressive rate limits.
★Replaced a paid ScrapeGraphAI news extraction pipeline with a custom Cheerio-based scraper that intelligently removes noise tags (scripts, footers, ads, sidebars), probes multiple container selectors, and validates articles against tracked tickers before insertion.
★Built a strategy backtester that simulates sentiment-driven trading strategies against historical database records with entry/exit condition limits, Monte Carlo fallback pricing, and comparative benchmark yield plots against passive hold.
★Implemented advanced quantitative tools: Lag Correlation analysis between sentiment and price across multiple time offsets, AI Narrative 2D projection using UMAP-like dimensionality reduction, Temporal Narrative Trajectory tracking via leader clustering on vector embeddings, and a Whale vs. Retail Crossover Index grounded in real database telemetry.
★Designed a premium dark UI with tactical aesthetic — animated SVG arc horizon masks, floating particle canvases, spring-animated sentiment gauge needles, retro CRT terminal loader with rotating crypto history facts, Obsidian-style guestbook dot grid, and hover-triggered typographic animations using Instrument Serif.
★Built the entire ingestion pipeline on cron-job.org with two scheduled jobs (30-min flash scraper, 6-hour Reddit batch), a scrape_cache table for duplicate prevention, and pre-computed 1-hour timeseries rollups with recency decay, source credibility weights, and log-scaled upvote boosts.