Demos
Use this page to choose a public AGILAB demo route. It is a router, not a quick-start guide.
Choose a demo
What each route is for
AGILAB demo: use AGILAB Demo for the self-serve public Hugging Face Spaces route for the AGILAB web UI. It publishes the lightweight
flight_telemetry_projectandweather_forecast_projectpaths, so use it as the public first proof forPROJECT->ORCHESTRATE->WORKFLOW->ANALYSIS, includingview_maps,view_forecast_analysis, andview_release_decision.agi-core demo: notebook-first runtime path. Use this if you want the smaller
AgiEnv/AGI.run(...)surface before the web UI.Notebook migration demo: use Notebook Migration Example when you want the notebook-to-AGILAB story: source notebooks, migrated
lab_stages.toml,pipeline_view.dot, exported forecast artifacts, and the hostedweather_forecast_projectanalysis route.PyTorch Playground: use PyTorch Playground when you want a browser-visible neural-network lesson that goes beyond a classic playground: live play/pause training, replayable lesson configs, PyTorch code handoff, and an evidence ZIP from the same run.
Advanced Proof Pack: use Advanced Proof Pack after the first demo when you want the deeper packaged proof routes:
mission_decision_project, theexecution_pandas_projectCython worker speedup demo,execution_polars_project, thesqlite_connector_proofdatabase evidence preview, UAV queue analysis withuav_relay_queue_project,service_modepreviews,inter_project_dagpreviews,mlflow_auto_trackingpreviews,resilience_failure_injectionpreviews,train_then_servepreviews, Data Connectors, and Release Proof.Industrial optimization examples: use Industrial Optimization Examples when your apps repository includes
sb3_trainer_projectand you want the advanced SB3 routes: Active Mesh Optimization, MLflow auto-tracking, multi-app DAGs, resilience/failure injection, and train-then-serve contracts.Quick start: the safest truthful first proof of the full product path. Use Quick-Start if you want the recommended local run instead of a public demo.
Short demo routes
Use these as narrow product demos. They are intentionally generic and should not depend on private apps or app-specific claims.
Robot/proof coverage
Every route on this page must be tied to one of the local validation contracts:
UI robot: browser-visible AGILAB routes are covered by
tools/ui_robot_coverage_contract.py --json. The contract checks the local all-built-inui-robot-matrixprofile, the hosted first-proof visual robot, the hosted install robot, and the configured apps-pages used byflight_telemetry_project,weather_forecast_project,mission_decision_project,execution_pandas_project,execution_polars_project,uav_queue_project, anduav_relay_queue_project.Static/CLI proof: non-browser preview routes are covered by
tools/public_proof_scenarios.py --compact. This includes the local package proof, hosted weather proof, MLflow proof, distributed-worker health proof, notebook migration proof, resilience failure-injection proof, train-then-serve proof, and service-mode preview proof.
The static scenario contract is available as JSON:
uv --preview-features extra-build-dependencies run python tools/public_proof_scenarios.py --compact
uv --preview-features extra-build-dependencies run python tools/public_proof_scenarios.py --first-proof-json first-proof.json --hf-smoke-json hf-space-smoke.json --output public-proof-scenarios.json
- Local app proof
Install the released examples profile or use the source checkout, then run the public first proof:
python -m pip install "agilab[examples]" python -m agilab.lab_run first-proof --json --max-seconds 60
Stop when the command exits successfully and writes
run_manifest.json. This package route usesagi-appsas the public app umbrella and resolves the built-in project from the matching per-app package. Installagilab[ui]and rerun with--with-uiwhen you also want to boot the packaged local pages andagi-pagesanalysis views. The same route is available in the UI by followingPROJECT->ORCHESTRATE->ANALYSISwithflight_telemetry_project.- PyTorch Playground teaching route
Use PyTorch Playground when the demo objective is visual ML learning plus reproducible engineering handoff:
agilab app surface pytorch_playground_project --ui streamlit
Stop when the boundary-first panel is visible, press
Run instant demoforInstant wow: clean circles, read the start/now replay, try the XOR lesson card, and download the evidence ZIP. The useful proof is not the visual boundary alone; it is the replay token, manifest, generated PyTorch/Lightning snippets, and boundary snapshots attached to that lesson.- SQLite database proof
Use the packaged
sqlite_connector_proofpreview when the demo objective is reproducible database access without a remote database, Docker, or secrets:uv --preview-features extra-build-dependencies run python src/agilab/examples/sqlite_connector_proof/preview_sqlite_connector_proof.py --output-dir /tmp/agilab-sqlite-proof
Stop when
sqlite_connector_proof.db,promotion_candidates.csv, anddatabase_evidence.jsonare visible. The evidence records the connector ID, SQLite driver, read-only query mode, schema hash, query hash, row count, result hash, and artifact hashes. This is the public first step before a Postgres, warehouse, or cloud SQL connector.- Cython worker speedup demo
Use Execution Playground when the demo objective is performance engineering rather than a domain story. Select
execution_pandas_project, keep the defaulttyped_numerickernel and Cython setting, runINSTALLthenEXECUTE, and inspect the reducer evidence forkernel_mode,kernel_runtime, anddtype_contract. The versioned local kernel proof records0.620sPython vs0.002sCython on 100,000 rows x 32 passes, with matching checksums and a306xhot-loop speedup.For a short command-line proof of the compiled hot loop, run:
uv --preview-features extra-build-dependencies run python tools/benchmark_execution_pandas_cython_kernel.py --rows 100000 --compute-passes 32 --repeats 3 --warmups 1
Stop when the Python and Cython checksums match and the report shows the Cython runtime separately. This is a kernel-scoped speedup demo; full AGILAB runs still include CSV reads, dataframe grouping, result writes, worker startup, and optional Dask/process orchestration.
For the real-world version of the same pattern, run
flight_telemetry_project. Its Polars ingestion and map/network analysis stay in Python, while the haversine speed kernel recordsspeed_kernel_runtime,speed_dtype_contract, and checksum evidence in the reducer summary.- Rust/PyO3 native-worker preview
Use the packaged
native_rust_workerpreview when the demo objective is to show AGILAB’s native-worker extension boundary without adding Rust to the base install:uv --preview-features extra-build-dependencies run python src/agilab/examples/native_rust_worker/preview_native_rust_worker.py --output-dir /tmp/agilab-rust-worker
Stop when
native_rust_worker_evidence.jsonand the generatedrust_workerPyO3/maturin skeleton are visible. This is an advanced worker extension preview: AGILAB orchestration and evidence stay in Python, only a measured typed hot kernel moves to Rust, and compiling the generated extension remains an explicit follow-up step.- Distributed worker route
Use the same public app, then switch ORCHESTRATE from the local path to the configured worker or SSH-host path. Keep the demo bounded: prove that worker packaging is staged, service health gates report status, and outputs land under the normal log directory.
uv --preview-features extra-build-dependencies run python tools/service_health_check.py --format json
Stop when the health gate is explicit. This is a worker/operator demo, not a certification of every possible remote topology.
- MLflow tracking route
Start with the packaged preview when you want a short, deterministic proof:
uv --preview-features extra-build-dependencies run python src/agilab/examples/mlflow_auto_tracking/preview_mlflow_auto_tracking.py --output-dir /tmp/mlflow_auto_tracking_preview
Add
--with mlflowto theuvcommand when you want the same evidence logged into a local MLflow store. The demo objective is to show that AGILAB keeps setup, execution, artifacts, and visible results together while MLflow remains the tracking system of record when it is used.Stop when the pipeline artifacts and the MLflow run link point to the same experiment evidence.
- Resilience failure-injection route
Use the packaged preview when the demo objective is strategy comparison under a controlled degradation event:
uv --preview-features extra-build-dependencies run python src/agilab/examples/resilience_failure_injection/preview_resilience_failure_injection.py --output /tmp/resilience_preview.json
Stop when the route ranking before failure, route ranking after failure, and recommended fixed/replanned/search/policy response are visible in the same JSON payload. The preview is deterministic and does not train a real policy.
- Train-then-serve route
Use the packaged preview when the demo objective is the handoff from experiment evidence to a service-ready contract:
uv --preview-features extra-build-dependencies run python src/agilab/examples/train_then_serve/preview_train_then_serve.py --output-dir /tmp/train_then_serve_preview
Stop when
service_contract.json,prediction_sample.json, andservice_health.jsonare visible. The preview is deterministic and does not start persistent workers.- Notebook migration route
Use the packaged migration example when the demo objective is notebook consolidation rather than execution speed:
uv --preview-features extra-build-dependencies run python src/agilab/examples/notebook_to_dask/preview_notebook_to_dask.py --output /tmp/notebook_to_dask_preview.json
Then open Notebook Migration Example or switch the hosted UI to
weather_forecast_projectand openview_forecast_analysis. Stop when the notebook source, migrated pipeline shape, exported artifacts, and reusable analysis view are visible together.
Demo naming
Keep the two public AGILAB demo lanes separate:
flight_telemetry_projectis the default hosted/newcomer demo. It is a lightweight data-generation path used to prove the core UI and local execution flow quickly.weather_forecast_projectis the second hosted demo. It is a lightweight notebook-migration path with source notebooks, forecast artifacts, and release-decision views.uav_relay_queue_projectis the UAV Relay Queue RL demo. It is the advanced domain scenario and should not be described as the default hosted app.mission_decision_projectand the execution playground apps are advanced proof routes. They are public built-in demos, but they should not replaceflight_telemetry_projectas the default hosted/newcomer app.