Flyte & Union.ai Tutorials

Tutorials and examples for building AI agents, ML pipelines, and data workflows with Flyte 2.

---

👉 Tonight's workshop: Fine-tuning open models with agents

tutorials/langgraph-model-factory-agent/

An ML engineer agent runs the model factory: evaluate, fine-tune, promote, deploy, and retrain when the data drifts. LangGraph for the agent, Flyte 2 for everything it touches.

A support agent in production routes tickets with an API call. An ML engineer agent gets the request to replace that call with a model we own: it evaluates open candidates in parallel, fine-tunes the ones worth it, promotes the winner, deploys it, and tests the live app. Then the data drifts and the loop turns again. Every tool call is a durable Flyte task, so a run that dies halfway resumes without retraining anything, and every eval and fine-tune leaves a report you can open.

Three ways to run it (Setup has the details):

• A Flyte devbox on your own machine (recommended: the full workshop). Docker Desktop, Python 3.10+, uv; flyte start devbox --apple-silicon on an Apple Silicon Mac, plain flyte start devbox with an NVIDIA GPU or on CPU.
• The workshop's Union cluster (tryv2), from any laptop. One flyte create config command and a browser sign-in; the cluster's GPUs do the work.
• The notebook on Google Colab (no GPU, nothing to install). Click the badge. Slightly pared down: the core loop end to end, without the live app and kill-and-resume steps.

Bring a laptop. The devbox and Colab paths need a Claude or OpenAI API key; the cluster already has one.

git clone https://github.com/unionai/workshops
cd workshops/tutorials/langgraph-model-factory-agent
uv venv .venv --python 3.11 && source .venv/bin/activate
uv pip install -r requirements.txt

flyte start devbox --apple-silicon && flyte create config --devbox      # a Mac: the devbox path
# or, the cluster path:
# flyte create config --endpoint tryv2.hosted.unionai.cloud --project factorysea --domain development --builder remote

flyte run model_factory.py model_factory --fine_tune_too                # step 1: the factory, you at the controls

---

Featured

Example	Description
Apple Silicon: fine-tune on your Mac's GPU	flyte start devbox --apple-silicon, gpu="AppleSilicon": LoRA on mps for text-to-SQL (5% → 62% exact match, 14 s of training), live loss chart, the adapter served on the same GPU; one env var moves it to a T4
Apple Silicon: serve an LLM on your Mac	Qwen3-1.7B behind an OpenAI-compatible Flyte app on the Metal GPU (streaming, tool calls), benchmarked from a task, torch on mps vs MLX native (79 vs 277 tok/s), a chat UI that finds the app by name
Apple Silicon: an agent whose model and tools run on your Mac	A LangGraph SQL agent with Qwen3-4B served on the Mac's GPU as its brain and Flyte tasks as its tools: every query is a durable child action, 16 questions fanned out and scored (81%), swap the brain for Claude with one env var
Model Factory Agent	Tonight's workshop. A LangGraph agent plays ML engineer on Flyte 2: parallel evals, fine-tunes, promotion, a deployed router app it tests, drift and the retrain; runs on a Union cluster, on a Flyte devbox on your Mac, or in Colab with no cluster at all
Model Factory Agent + Grafana	A LangGraph agent plays ML engineer: evaluates five candidates on T4s in parallel, fine-tunes the right ones (and discovers a 149M encoder beats the chat models), promotes it as a Union artifact, deploys it, tests the live app, and turns the factory again when the data changes; Grafana Agent Observability watches every turn
SkyRL SQL Agent	Multi-turn RL against a real database — write one SkyRL-Gym environment, then drive it with no model, with a 0.5B, with Claude, or with SkyRL's distributed trainer; GRPO rollouts fan out as durable Flyte tasks
RAG and Agentic Memory	One vector store, pointed two directions — build a RAG index, watch retrieval work with no model involved, visualize the embedding space, then let an agent write its own memories back into it
Code Mode — NYC Taxi analyst	Claude writes one program, the Monty sandbox runs it, and its loops fan out into durable parallel tasks over real NYC taxi data
LangGraph Research Pipeline	Research agent pipeline — LangGraph orchestrates planning and quality gates, Flyte fans out parallel researcher tasks
LangGraph ReAct Agent	Build a ReAct agent with LangGraph + OpenAI on Flyte
Stable Diffusion	Generate images from text prompts with SDXL Turbo on GPU
DuckDB ETL	Extract and transform data with DuckDB SQL
Fraud Detection with Feast	Real-time fraud scoring with Feast feature store + XGBoost

---

Get Started

Tutorial	What you'll learn
Flyte Basics	Flyte 2 fundamentals — tasks, pipelines, error handling, TaskEnvironment, ReusePolicy, map()
LangGraph ReAct Agent	Build a ReAct agent with tools in a single file
Stable Diffusion	GPU inference with Flyte reports
Image Classifier	Fine-tune ResNet18 on HuggingFace dataset with PyTorch
DuckDB ETL	Data pipeline with DuckDB SQL and Flyte reports
Snowflake ETL	ETL pipeline with the Snowflake connector
Flyte Local Dev	Local dev features — TUI, caching, reports, tracing, serving (no cluster needed)
Fraud Detection with Feast	Fraud scoring pipeline with Feast feature store, XGBoost, and real-time serving

---

Agents

Tutorial	Description
Apple Silicon SQL Agent	LangGraph ReAct agent on a model served from your Mac's GPU, tools as Flyte tasks, no API key; 16 scored questions, brain swappable to Claude or a cluster GPU
Model Factory Agent + Grafana	The agent runs the model factory: parallel T4 evals, fine-tunes, artifact promotion, a deployed router app it tests and can roll back, an artifact trigger that validates every promotion, a human approval gate, and a day-two retrain; observed in Grafana Agent Observability, with a crash-and-resume that trains nothing twice
Code Mode — NYC Taxi analyst	The agent writes a program instead of calling tools one at a time. It runs in the Monty sandbox, and a loop in the generated code becomes a fan-out of durable, parallel query tasks over 3M+ real taxi trips
RAG and Agentic Memory	One Chroma store, pointed two directions — build a document index, search it with no model, answer from it with citations, see the embedding space in 2D, then let an agent write its own memories back into it
LangGraph Research Pipeline	Research agent pipeline — LangGraph orchestrates planning and quality gates, Flyte fans out parallel researcher tasks via Tavily web search
LangGraph ReAct Agent	Single-file ReAct agent with LangGraph
Planner Multi-Agent System	Scalable planner multi-agent system
ReAct Multi-Agent System	Adaptive ReAct multi-agent system
Debate Multi-Agent System	Multi-agent debate pattern
Manager Multi-Agent System	Manager-worker agent delegation
Sequential Multi-Agent System	Sequential agent pipeline
Autoresearch	Autoresearch-style self-healing agent on Flyte

MCP

Tutorial	Description
MCP Recipe Assistant	Build and deploy a recipe assistant MCP server on Union

LLM Fine-Tuning

Tutorial	Description
Apple Silicon: LoRA on mps	Fine-tune Qwen2.5-0.5B for text-to-SQL on your Mac's GPU from a Flyte devbox task, then serve the adapter; the same pipeline runs on a T4 with ACCELERATOR=T4
LoRA / QLoRA / Full	Fine-tune an LLM on text-to-SQL with LoRA, QLoRA, or full fine-tuning — live training reports, FastAPI serving, Gradio UI
Nemotron 3 Nano — LoRA / QLoRA	Fine-tune NVIDIA's Nemotron 3 Nano 4B for tool calling — LoRA on an L40S or QLoRA on a T4 from the same pipeline, discovered LoRA targets for the Mamba hybrid, adapter serving
Nemotron 3 Nano — GRPO	Reinforcement learning on Nemotron 3 Nano 4B — Countdown puzzles with a verifier as the only reward, thinking off, live solve-rate charts
Nemotron — Healthcare: clinical notes	Fine-tune Nemotron 3 Nano 4B to turn clinical case notes into fixed-schema JSON records — field-level scoring, LoRA on an L40S or QLoRA on a T4
Nemotron — Finance: XBRL tagging	Fine-tune Nemotron 3 Nano 4B to tag amounts in SEC filings with US-GAAP XBRL concepts (FiNER-139) — exact tag-set and pair-F1 scoring, LoRA on an L40S or QLoRA on a T4
Nemotron — Legal: contract clauses	QLoRA an 8B model on a T4 — Llama-3.1-Nemotron-Nano-8B classifies contract clauses into the 100-way LEDGAR taxonomy, reasoning switched off via system prompt
Nemotron — Biomed: chemical-protein relations	Fine-tune Nemotron 3 Nano 4B to classify chemical-protein relations in PubMed abstracts (ChemProt) — the extraction step behind drug-target knowledge bases, LoRA on an L40S or QLoRA on a T4
GRPO — Code Generation	Teach a model to write Python with GRPO — reward = sandboxed test execution, MBPP dataset, live reward/pass-rate charts
GRPO — Distributed	Scale GRPO across the cluster — fan out sandboxed verification to a reusable pool, then disaggregate rollouts onto vLLM workers with LoRA weight sync
GRPO	GRPO fine-tuning on math/reasoning tasks
DPO	Direct Preference Optimization for alignment
PPO	Proximal Policy Optimization for RLHF

ML / AI

Tutorial	Description
Fraud Detection with Feast	Fraud scoring pipeline with Feast feature store, XGBoost, and real-time serving
Stable Diffusion	Image generation with SDXL Turbo
Image Classifier	Fine-tune ResNet18 on Beans dataset
DETR Object Detection	Fine-tune DETR for object detection with live mAP charts

Biotech / Life Sciences

Tutorial	Description
Genomic Variant Effect Prediction	Score DNA mutations with HuggingFace Carbon genomic foundation model — zero-shot pathogenicity prediction on BRCA2, TP53, KRAS, and more
DNA Sequence Generation & Analysis	Generate DNA with Carbon and compare to real genes — GC content, codon usage, ORFs, dinucleotide frequencies
Gene Comparison Across Species	Compare homologous genes across 6 species with Carbon scoring, phylogenetic trees, and ESMFold 3D structure comparison
Protein Sequence Analysis	Analyze protein properties, compute sequence similarity, run ESM-2 embeddings, and predict 3D structures with ESMFold
Drug Molecule Screening	Virtual drug screening — compute physicochemical properties, apply Lipinski's Rule of Five, rank candidates by drug-likeness
Cell Microscopy Classification	Fine-tune a Vision Transformer (ViT) to classify blood cell types from microscopy images

Data

Tutorial	Description
DuckDB ETL	SQL-based data pipeline with DuckDB
Snowflake ETL	ETL with Snowflake connector
Lance Streaming for Vision	Convert a swarm of tiny per-sample image files (real CPPE-5 detection data) into one Lance dataset, benchmark per-file vs Lance streaming against real object storage, then train / evaluate / explore a Faster R-CNN streamed straight from S3 on a T4

---

Setup

# Clone the repository
git clone https://github.com/unionai/workshops
cd workshops

# Navigate to any tutorial
cd tutorials/langgraph_agent_research

# Create virtual environment and install dependencies
uv venv .venv --python 3.11
source .venv/bin/activate
uv pip install -r requirements.txt

Flyte Cluster (for remote runs)

flyte create config \
    --endpoint <your-endpoint> \
    --auth-type headless \
    --builder remote \
    --domain development \
    --project flytesnacks

Don't have a cluster? Request access at flyte.org.

Run examples

Remote:

uv run flyte run workflow.py research_pipeline --query "Compare quantum computing approaches"

Local:

uv run flyte run --local workflow.py research_pipeline --query "Compare quantum computing approaches"

Local with TUI:

uv run flyte run --local --tui workflow.py research_pipeline --query "Compare quantum computing approaches"

Start TUI dashboard:

uv run flyte start tui