A coding agent a whole team shares. The agent's transcript is stored against the project rather than against a person, and the project's standing instructions live in a COFOUND.md the agent re-reads on every run, so a second person's message continues the same conversation instead of starting a new one. Runs stream live to everyone in the project, writes can be gated behind teammate review, and a review comment rides into the next run's prompt as an instruction.

• TypeScript
• Next.js
• Anthropic API
• Redis
• Liveblocks
• GitHub OAuth
• Playwright

A production multi-agent RAG platform that answers questions about uploaded documents with inline citations, then grades its own answers. Five agents handle ingestion, query routing, hybrid retrieval (pgvector dense search fused with BM25 and reranked by a cross-encoder), streaming synthesis, and an automated RAGAS evaluation loop that scores every answer for faithfulness, context precision, and answer relevancy. Runs entirely on free-tier infrastructure.

• TypeScript
• Next.js
• pgvector
• LangChain
• RAGAS
• Groq
• HuggingFace
• Neon

A live quantitative research platform that classifies the US economy into four macro regimes (Expansion, Late-Cycle, Recovery, Contraction) using a Hidden Markov Model trained on 35 years of FRED data. A walk-forward validated AutoGluon ensemble forecasts next-month regime transitions. Full stack: Python HMM + FastAPI backend, Next.js 16 frontend, deployed on Vercel.

• Python
• hmmlearn
• AutoGluon
• FastAPI
• Next.js
• FRED API
• GSAP

An equity research notebook where every stock pitch is timestamped in git history and tracked live against the S&P 500, with no retroactive edits. Each pitch is a structured research note: Situation, Complication, and falsifiable thesis pillars, MECE risk tables with mitigants, comparable multiples frozen at pitch date, bear, base, and bull targets, and an interactive DCF with a WACC by terminal-growth sensitivity grid. A form-based Pitch Builder generates new entries.

• TypeScript
• Next.js
• Valuation
• DCF
• Equity Research

An earnings call intelligence platform that scores SEC EDGAR transcripts with FinBERT sentiment analysis, then adjusts each score for the prevailing macro regime supplied by MacroScope. The result is a sentiment surprise signal: whether a company sounded better or worse than the economic climate would predict. Analyzes the latest earnings release for any ticker live, backed by a 38-call historical dataset across 12 S&P 500 companies, with generated PDF reports.

• Next.js
• TypeScript
• FinBERT
• HuggingFace
• SEC EDGAR
• Groq

A BERT-based sentiment classifier fine-tuned on financial news headlines to flag positive, neutral, and negative market sentiment. Four epochs of fine-tuning with AdamW and a linear learning-rate schedule pushed accuracy from 70% to 99%, turning a generic language model into a domain-specific market-sentiment signal.

• Python
• PyTorch
• BERT
• HuggingFace
• Transformers
• NLP

Analysis of 260,000 California traffic collisions (SWITRS, 2006–2021) to predict crash severity. After preprocessing and addressing class imbalance with SMOTE, a Random Forest model reached 71% weighted accuracy, outperforming KNN and linear baselines, with feature importance analysis surfacing the conditions most associated with severe outcomes.

• Python
• Scikit-learn
• Pandas
• AutoGluon
• SMOTE

A computational pipeline that generates rocket fin geometries in OpenSCAD, runs them through 1,000+ OpenFOAM simulations, and trains an AutoGluon ensemble to predict drag from geometric parameters. SHAP analysis reveals which fin dimensions matter most. The ensemble reached R² = 0.989, outperforming polynomial regression and standalone random forests.

• Python
• AutoGluon
• SHAP
• OpenFOAM
• OpenSCAD