Real buy-side adoption, verified institutional cases, and the golden rule: the LLM reasons, the math decides, the human vetoes.
// open course
QUANT AGENTIC ORCHESTRATION
Governed LLM agents for hedge funds, risk and portfolio management. A free, open course that takes you from LangChain v1 fundamentals to designing a governed multi-agent hedge fund, with real code and verified institutional sources.
| Parts | 12 modules + 3 appendices |
|---|---|
| Sources | 421 verified |
| Stack | Python · LangChain v1 |
| Level | Intermediate-advanced |
| Price | Free, no sign-up |
Index
- 01The LLM stack in asset management in 2026
- 02LangChain v1 essentials for quants
The Runnable mental model, LCEL, structured output with Pydantic, and low-level tool calling.
- 03Agents and LangGraph: orchestration for investment decisions
create_agent, middleware, StateGraph, persistence, real human-in-the-loop, and multi-agent patterns.
- 04Market and documentary data for LLM systems
2026 source map, programmatic SEC EDGAR, news and sentiment, and point-in-time pipelines without traps.
- 05Institutional-grade financial RAG
Structure-aware chunking, hybrid search with reranking, the text/numbers split, and RAGAS evaluation.
- 06Prompt engineering and structured output for finance
Role-based system prompts, few-shot, judicious chain-of-thought, and anti-numeric-hallucination grounding.
- 07Quantitative risk: from VaR to the risk copilot
Exact VaR and Expected Shortfall, stress testing, and the LLM as an explanation layer, never a calculator.
- 08Portfolio construction: from Markowitz to LLM-generated views
Efficient frontier, Black-Litterman with LLM views, and optimizer sensitivity.
- 09LLM signals and rigorous backtesting
From text to signal, backtests without look-ahead, realistic costs, and the deflated Sharpe ratio.
- 10Governance, compliance, and regulation of agentic AI
AI Act, NIST AI RMF, model risk management, and auditable evidence for agentic systems.
- 11Production engineering for financial LLM systems
Observability with LangSmith, continuous evaluation, costs, latency, and governed deployment.
- 12Capstone project: a governed multi-agent hedge fund
Full integration: research, risk, portfolio, and execution with human sign-off in a single system.
- AAppendix A · Quick-reference formula sheet
Every formula in the course, checked against the primary literature, in a single reference.
- BAppendix B · Lab environment
Full environment setup: dependencies, APIs, paper trading, and lab conventions.
- CAppendix C · Glossary
Course terminology: from LCEL to point-in-time, precisely defined.
- References · 421 verified sources
All the course evidence: papers, official docs, engineering blogs, and primary sources, verified as of July 2026.
What this is and what it is for
Major funds no longer ask whether to use language models, but how to govern them. Kensho (S&P Global), Bloomberg, JPMorgan, BlackRock and Man Group run agentic systems in production today. This course teaches exactly that: how those systems are built, evaluated and governed when real money is at stake.
What you will learn
- Orchestrate LLM agents with LangChain v1 and LangGraph: tools, middleware, persistence and human control.
- Build institutional-grade RAG over filings, earnings calls and regulatory texts.
- Apply LLMs to risk (VaR/ES), portfolios (Black-Litterman) and signals with rigorous backtesting.
- Govern it all: compliance, AI Act, observability and production.
How to use it
- No sign-up, no accounts. Read in order or jump to the module you need; each one is self-contained.
- Every module ends with an interactive understanding check and exercises in code and design.
- All code is copyable and runnable in the lab environment of Appendix B.
- Free pace: ~20–40 min of reading per module, plus practice.
Who it is for
- Quant developers moving into agentic LLM systems.
- Risk managers and PMs who need to understand what these systems can (and cannot) do.
- Anyone with intermediate Python (pandas, numpy) and quantitative finance fundamentals.
- New to LLMs? Module 1 and the «EXPANDE» boxes bring you up to speed without unnecessary jargon.
The principle running through all 15 parts
The LLM reasons · the math decides · the human vetoes
No order goes out without a human signature. No number is invented by the model: a deterministic tool computes it and the LLM explains it. Every module applies this boundary — understand, reason, act — backed by the institutional evidence behind it.
Working material, not lecture notes
Real code
LCEL pipelines, agents with middleware, audited tool calling, bull/bear debates with LangGraph, risk dashboards: copyable, commented LangChain 1.x code aimed at production.
Evidence, not hype
421 verified sources: documented deployments at Kensho, Bloomberg, JPMorgan, BlackRock, Man Group, Bridgewater, AQR, Balyasny and Point72 — and also the independent replications that debunk the hype.
Quantitative rigor
Exact formulas checked against the primary literature, point-in-time data, realistic costs and the deflated Sharpe ratio. Where there is no evidence, the course says there is none.