Build It Yourself
Eight labs, one project. You build a Financial Intelligence Assistant layer by layer: a RAG system, then an agent, then guardrails, then full observability.
The Capstone
The labs are not isolated exercises. Each one extends the previous build, so by Lab 7 you own a RAG-powered, agent-augmented, guardrailed and observable system for financial document analysis.
Labs 0, 1, 2 and 5 feed the capstone (clean text, model choice, prompts, governance documentation) and can be done on their own.
Document Preprocessor
Module 0 · NLP Foundations
NLP pipeline on a real regulatory circular: cleaning, tokenisation, NER, TF-IDF.
Model Comparison Bench
Module 1 · Foundation Models
Run one financial task across Claude, GPT-4o and Gemini; compare accuracy, tone and variance at temperature 0 vs 0.7.
Financial Prompt Library
Module 2 · Prompt Engineering
Five production-ready prompt templates, zero-shot vs few-shot vs CoT, and prompt caching cost measurement.
Regulatory Document RAG
Module 3 · Embeddings & Vector Stores
Chunk and embed three public documents into ChromaDB, then answer questions with citations. Start here.
Financial Research Agent
Module 4 · AI Agents
Add three tools and a ReAct loop on top of Lab 3, with a human-in-the-loop checkpoint for high-stakes answers.
Bias Audit & Model Card
Module 5 · Governance
Audit a mock credit-scoring dataset, compute fairness metrics and SHAP values, then write a model card.
Security Guardrail Layer
Module 6 · AI & GenAI Security
Harden the Lab 4 agent: Presidio PII detection, spotlighting, LLM Guard injection checks, and a Garak attack run.
Full Observability Stack
Module 8 · Observability & Evals
Trace the agent in Langfuse, build a 20-question golden set, and report groundedness, latency and cost per query.
Choose Your LLM Provider
Every lab is written so the model provider is a single setting. Pick whichever fits your budget and your data-privacy needs, and the rest of the code stays the same.
- ▸Runs open models on your own machine, so no data leaves it
- ▸No API key and no rate limits
- ▸Needs a decent laptop (8 GB+ RAM) and a model download
- ▸Best for privacy-sensitive financial data
- ▸Fast hosted open models via an API key
- ▸Free tier is rate-limited; check current limits
- ▸OpenAI-compatible API, so setup is easy on Colab
- ▸Best when your laptop is too small for Ollama
- ▸Strongest results for long documents and tool use
- ▸Needs an API key with billing enabled
- ▸Costs for the whole lab series are small but not zero
- ▸Best for the closest match to the module examples
Lab 4 adds three more options: Google Gemini (free tier), OpenRouter (free models with small limits) and OpenAI (paid). Embeddings in Labs 3 and 4 run locally with a small open model through ChromaDB, so they need no key on any path. Lab 1 compares several providers by design, so it uses more than one. Results and quality will differ between providers; expected outputs on each lab page note where that matters.
Before You Start
- ▸Accounts: a Google account for Colab, plus access to one LLM provider from the table below. You do not need a paid API key to complete the labs.
- ▸Order: read the matching module first, then do the lab. Labs 3, 4, 6 and 7 build on one another, so do them in sequence.
- ▸Data: every lab uses publicly available documents, with sources listed on each lab page.