AI Lab

A personal lab for developing and testing focused patterns, techniques, and subsystems for agent and harness development.

Some projects are small learning exercises; others are more involved framework experiments. They are grounded in practical agent-system concerns, but built here as personal pattern studies and experiments that others can learn from too.

Most examples live under src/ as standalone workspace packages. A common way to run one is:

uv run --package <package-name> python src/<folder>/main.py

Index

PydanticAI Basics

• src/pydantic-ai-agent-basic: smallest possible pydantic-ai agent example.
• src/pydantic-ai-output-basics: output patterns and structured response basics.
• src/pydantic-ai-agent-tools: tool-using agent with calculator, search, and weather.

Human-In-The-Loop

• src/pydantic-ai-human-approval: CLI deferred-tool approval (orders and support cases).

Data Analysis

• src/pydantic-ai-query-dataset: agent that answers questions by running DuckDB SQL over a dataset.

Research Workflows

• src/pydantic-ai-deep-researcher: graph-driven iterative web researcher that plans, executes, and adapts across multiple cycles.

History And Memory

• src/pydantic-ai-history-summarizer: summarize conversation history for reuse.
• src/pydantic-ai-history-trimmer: trim history while keeping useful context.
• src/history-compaction-framework: reusable framework for projected history compaction.
• src/memory-framework: reusable file-backed topic memory framework.

Observability

• src/pydantic-ai-logfire: Logfire instrumentation example for pydantic-ai.

MCP

• src/hello-world-mcp: minimal MCP server example.

Embeddings And Audio

• src/embedding-foundations: introductory embedding notebook/package scaffold.
• src/text-to-speech: text-to-speech example.

Applied ML Pipelines

• src/gb-churn: staged synthetic churn pipeline with SQL-style transformations.
• src/gb-340b: staged synthetic 340B CE audit prioritization pipeline.

Deeper Notes

Human-In-The-Loop

This section is about agent flows where the model can chat and clarify normally, but a human must explicitly approve side-effectful tool execution.

Examples:

• src/pydantic-ai-human-approval: CLI deferred-tool approval around order and support-case submission.

Focus:

• a chat-first workflow where the agent gathers missing details naturally
• a hard approval gate only at the tool boundary, not at every reply
• direct review of validated deferred tool args from DeferredToolRequests
• resuming the same run with DeferredToolResults after yes, revise, or cancel

Data Analysis

src/pydantic-ai-query-dataset is less about generic agent basics and more about a specific analysis pattern: let the model translate user questions into SQL over a constrained dataset, then answer from the query results.

Examples:

• src/pydantic-ai-query-dataset: natural-language questions answered through DuckDB SQL over a dataset.

Focus:

• natural-language-to-SQL over a known dataset
• agent-guided analysis without giving the model arbitrary code execution
• a useful bridge between general tool use and more structured analytical workflows

History And Memory

These examples are about managing long-running context, but they do it at different layers: trimming the live conversation, compacting it more systematically, or storing reusable memory outside the immediate transcript.

Examples:

• src/pydantic-ai-history-summarizer: replace raw prior turns with a compact summary.
• src/pydantic-ai-history-trimmer: keep recent or important turns while pruning lower-value history.
• src/history-compaction-framework: reusable framework for projected history compaction.
• src/memory-framework: file-backed persistent memory outside the live conversation history.

Focus:

• summarize or trim when the problem is mostly token pressure
• use a compaction framework when you want that behavior to be reusable and policy-driven
• use persistent memory when the problem is cross-session recall, not just context window size

Applied ML Pipelines

These examples are less about agents and more about structured ML workflows with staged, inspectable transformations over tabular data.

Examples:

• src/gb-churn: staged churn prediction pipeline for a classical tabular business problem.
• src/gb-340b: staged prioritization pipeline for a synthetic healthcare/compliance-style problem.

Focus:

• examples of non-agent AI work that is still highly practical
• pipelines where SQL-style transformation thinking is part of the design
• a contrast with the pydantic-ai examples, which are focused on agent/runtime patterns rather than classical model pipelines