Senior AI Engineer building agentic workflows: LLM-powered agents that plan, use tools and carry out multi-step work. My background is in systematic macro investing and machine learning for financial markets.
- Agentic workflows: multi-step agents, tool use, and orchestration
- Applied machine learning: data scrubbing, ensemble modeling, XGBoost, ML for forecasting
- Quantitative finance: signal research, factor modeling, portfolio construction, and backtesting
- Using Agentic RAG to Identify Price Chart Patterns: three Claude agents annotate SPY's 10-year and 3-year charts with support and resistance, trendlines and chart patterns (double tops, head and shoulders, triangles, wedges), then give a bullish or bearish verdict. Each annotation is grounded in definitions retrieved from a BM25 corpus built from StockCharts ChartSchool and ThePatternSite. Code finds every price level and pattern first. A Researcher retrieves cited definitions, a Chart Analyst picks what to draw and describes it, and a Verifier re-checks each item against the price data and the cited text. Code guardrails require every citation to have been retrieved, every quote to be verbatim and every dollar figure to match the data; Claude Code hooks block deletes and writes outside the project. Token costs are logged per agent and tool (about $0.68 a run). Evals: 100% retrieval recall@5, 4/4 planted patterns found with the correct verdict, a prompt-injection test passed, and 88 offline tests. The Verifier also caught two detector bugs. Includes a 7-page PDF.
- Study Sprint MCP: the learning app rebuilt on the Model Context Protocol. The five agents reach their tools through an MCP server, with a capability token per agent, tool pinning against poisoning, and a guardrail hook pipeline. A read-only learner server serves the lessons to Claude Desktop and Claude Code. A cost ledger prices every tool in tokens and dollars, and MCP itself is measured too: 0 extra tokens and about 1.8 ms per call. Evals: 27/27 live lessons over MCP, 67/67 MCP red-team cases, 216/216 direct-vs-MCP parity, and 14 planted defects caught. Includes a 15-page PDF of charts and commentary.
- Replicating an Asset Management Firm with Agents and Task Decomposition: five Claude agents replicate a firm's workflow, from data operations through quant research, performance analytics and investor relations to the investment team. They backtest a moving-average trend strategy and a five-lookback multi-asset momentum strategy (50/30/20 in the top three ETFs), compare both with SPY and QQQ, and produce a 29-page investor PDF. Code guardrails verify every number in it. Evals: 14/14 scorers in 3 of 3 live trials, 43/43 red-team cases and an independent cross-check.
- Agentic Prototype for an Encrypted Personal Assistant: a coordinator delegates to security, travel and household agents. They check passwords against Have I Been Pwned using k-anonymity, search flights, and track groceries and supplies. Everything is stored in an AES-256-GCM vault, and secrets never reach the model. Evals: 30/30 live conversations with no secret in any request, 72/72 red-team cases, and an independent cross-check of the supply math.
- OpenTable for Banquet Halls Using Agentic Workflows: an OpenTable-style booking app for banquet halls. A concierge agent delegates to venue-scout and pricing agents. Agents can only place holds; confirming a booking takes a human. Evals: 27/27 live conversations, 66/66 red-team cases, and 4,816 quotes cross-checked against an independent pricer. The live trials caught a real concurrency bug, now fixed.
- Dynamic Interactive Learning App Using Agentic Workflows: pick any subject and grade (K to College), and five Claude agents build a lesson with three key concepts, an SVG diagram and a 5-question quiz. Intake scopes the topic, Curriculum writes the concepts, Diagram and Quiz work in parallel, and a Reviewer checks the facts. A 1-minute timer triggers random movement breaks (jumping jacks, high knees, push-ups). Code measures reading level, sanitizes diagrams, blocks unsafe subjects and grades on the server. Evals include 185 tests (with Playwright browser tests), 45 red-team cases and a live K-to-College matrix (9/9). Includes a 12-page PDF guide.
- Quant-Trading-Templates: eight end-to-end quant projects covering momentum, breakouts, smart beta, alpha factors, NLP on 10-Ks, neural sentiment, signal combination and backtesting
Python · Claude / Anthropic API · MCP (Model Context Protocol) · PyTorch · TensorFlow · scikit-learn · pandas · FastAPI · AWS