Prerequisites
Comfort reading and writing Python (NumPy / pandas), working knowledge of probability and statistics, and core finance — portfolio mathematics, time value of money and basic derivatives.
✓ Advanced in finance and Python; no prior LLM or promptengineering experience assumed
✓ Optimization and stochastic processes help on the trading and asset-management tracks
Schedule at a glance
Session 1. Foundations and the Claude–GPT benchmark
Tues Oct 6 – 12:00 – 15.00 New York Time
Session 2. Guarded finance workflows
Wed Oct 10 – 12:00 – 15.00 New York Time
Session 3. Build the verifier-backed Skill
Thurs Oct 11 – 12:00 – 15.00 New York Time
Session 4. Optional debugging clinic, plus the between week field test
Fri Oct 12 – 12:00–13:00 New York Time
Session 5. Build the bounded finance agent
Tues Oct 13 – 12:00 – 15.00 New York Time
Session 6. Evaluate, attack and present
Thurs Oct 15 – 12:00 – 15.00 New York Time
15.0 Hours – Three artifacts + Twelve reference implementations
The Curriculum: 12:00 – 15.00 New York Time
Tues Oct 6 · 3 hours – 12:00 – 15.00 New York Time
Session 1. Foundations and the Claude–GPT benchmark
You ship: a configured dual-model workspace and a benchmark card for your own function
Topics: What frontier models are, what they are not, and the four failure modes that matter most in finance: fabricated numerics, stale data, invented citations, overconfident conclusions. The four levers of specialisation — Skils, tool use, retrieval and disciplined context — and why fine-tuning is rarely the first move. Benchmark discipline: pin the task, data, toolset, output schema and budget before comparing the two models, and keep answer quality separate from latency, cost, refusal behaviour and disagreement.
Lab: Configure the course workspace, run one function-specific task through both models, verify the numbers in code, inspect sources and structured output, and record where the models agree, disagree, or require a human handback.
Deliverable: A dual-model workspace, one reproducible benchmark run, and a short decision rule for which model — or which human review — a given task requires.
Wed Oct 10 · 3 hours – 12:00 – 15.00 New York Time
Session 2. Guarded finance workflows
You ship: one reproducible workflow with numerical, source and budget controls
Topics: When the right object is a deterministic script, an AI-augmented workflow or a bounded agent. Tool ca ling, structured output, current-data checks, source citation, explicit assumptions, cost limits and fail-closed behaviour. Derived values are computed in code; source-reported values stay attached to their source. Trading-track controls are built in: frozen evaluation windows, transaction costs, walk-forward checks, and an explicit warning once many variants have been searched.
Lab: Adapt the guarded reference workflow for your chosen function, then run a second function end to end — the fastest way to see that the controls, not the domain, are what transfer. Inject a stale price, a missing source and a malformed numeric output, and confirm the workflow rejects or hands back rather than improvising.
Deliverable: A function-specific workflow with a declared data boundary, numerical verifier, citation record, cost budget and human hand-back rule.
Thurs Oct 11 · 3 hours – 12:00 – 15.00 New York Time
Session 3. Build the verifier-backed Skill
You ship: a tested Skill for your function, plus the evidence from its independent recomputation
Topics: What a Skil is, what a SKILL.md file contains, and when a Skil beats prompting or retrieval. Encoding method knowledge, notation, house style, input contracts and failure modes. The verifier pattern: independently recompute the key result and surface disagreement instead of returning one trusted number. Composability, tests, versioning, model portability and sharing across a team.
Lab: A verifier-backed Skil, unit tests, a model comparison note and a short operating guide, contributed to the shared cohort library.
Deliverable: A verifier-backed Skil, unit tests, a model comparison note and a short operating guide, contributed to the shared cohort library.
BETWEEN THE WEEKS — THE FIELD TEST
- RUN – the Skil on a supplied unseen case
- RECORD – one failure or one model disagreement
- PATCH – the method, or its declared boundary
- SUBMIT – the test report before session 4
This is the reason the course runs over two weeks rather than three consecutive days. A method that has never failed in front of you is not yet a method you can trust with an agent behind it.
Fri Oct 12 · 3 hours – 12:00–13:00 New York Time
Session 4. Optional debugging clinic, plus the between week field test
Tues Oct 13 · 3 hours – 12:00 – 15.00 New York Time
Session 5. Build the bounded finance agent
You ship: a corpus-grounded agent prototype with declared permissions and stop conditions
Topics: When retrieval beats long context or tools. Corpus design, chunking, citation precision, freshness, and the difference between a sourced claim and a plausible completion. Every retrieved document is untrusted input. Writing the harness specification before the agent runs: tool whitelist, corpus boundary, token and cost budget, stopping rule and human hand-back point — the same discipline the AI Agents in Finance Certificate develops in depth.
Lab: Build your functionʼs agent on a sma l supplied corpus, connect the Skil from week one, require page-resolving citations, and confirm the agent refuses work outside its declared boundary.
Deliverable: A running bounded agent prototype, its corpus configuration, permissions file, stop conditions, and a trace of the inputs, tools, sources and policy checks behind one decision.
Thurs Oct 15 · 3 hours – 12:00 – 15.00 New York Time
Session 6. Evaluate, attack and present
You ship: an evaluation and red-team report, plus the capstone prototype
Topics: Evaluation for correctness, citation support, calibration, refusal quality, cost, latency and ha lucination rate. Crossmodel dispersion: where Claude and GPT fail differently under one contract. Prompt injection, document poisoning, stale data, malformed tools and budget exhaustion. The model-risk note a reviewer needs in order to assess a controled pilot.
Lab: Run the evaluation harness against your agent, attack a different functionʼs agent, plant an adversarial instruction inside a document and verify that it is reported rather than obeyed. Present the result, including the failures and the unresolved risks.
Deliverable: An evaluated capstone prototype, benchmark card, red-team report, monitoring checklist and a reviewer-facing limitations note.
Download Event PDF