Local runnable examples for AkôFlow, with the same activity DAGs used in Desktop and in the project documentation.
The catalog contains 27 small CPU workloads, the full 58-activity Montage 050d scientific workflow, three OpenOil ensembles with 8, 16 and 32 members, an eleven-activity USGS Earthquake workflow, a five-activity SRA Search workflow, a fourteen-activity E. coli Variant Calling workflow, and a twelve-activity Orcasound hydrophone-audio workflow processing complete real observations. Each example includes input data, a pinned image carrying that data, or a verified acquisition manifest, a preparation command, a local runner, a validator, and local daemon infrastructure. Current Desktop captures are in each bundle's screenshots/current/, with their exact run and image identified in its manifest.
Install Go and Python 3.12+. Blender, Earthquake, SRA Search, Variant Calling, Orcasound, Montage and OpenOil also require Docker; Montage requires linux/amd64 support. Preparation downloads the Go YAML module and the pinned packages required by that example. Workloads then run on local input data. The small examples perform real computation; their teaching profiles do not require cloud credentials, hosted LLMs, or GPU hardware.
git clone https://github.com/UFFeScience/akoflow-examples.git
cd akoflow-examplesEach command below starts from the repository root and keeps your shell in that directory:
| Example | Local command (from the repository root) |
|---|---|
| OpenDroneMap split/merge | (cd scientific-workflows/opendronemap && ./prepare.sh && ./run.sh) |
| USGS Earthquake | (cd scientific-workflows/earthquake && ./prepare.sh && ./run.sh) |
| SRA Search — crAssphage Screening | (cd scientific-workflows/sra-search && ./prepare.sh && ./run.sh) |
| Variant Calling — E. coli Genome Comparison | (cd scientific-workflows/variant-calling && ./prepare.sh && ./run.sh) |
| Orcasound — Hydrophone Audio and Orca Call Candidates | (cd scientific-workflows/orcasound && ./prepare.sh && ./run.sh) |
| SOMOSPIE — Soil Moisture Inference | (cd scientific-workflows/somospie && ./prepare.sh && ./run.sh) |
| RNA-seq — Transcript Abundance and Quality Control | (cd scientific-workflows/rnaseq && ./prepare.sh && ./run.sh) |
| Sarek — Regional Germline Variant Calling | (cd scientific-workflows/sarek && ./prepare.sh && ./run.sh) |
| PyCBC / LIGO — Gravitational-Wave Matched Filtering | (cd scientific-workflows/pycbc && ./prepare.sh && ./run.sh) |
| Montage 050 — Astronomical Image Mosaicking | (cd scientific-workflows/montage-050 && ./prepare.sh && ./run.sh) |
| OpenOil ensemble — 8 members | (cd scientific-workflows/openoil-ensemble-08 && ./prepare.sh && ./run.sh) |
| OpenOil ensemble — 16 members | (cd scientific-workflows/openoil-ensemble-16 && ./prepare.sh && ./run.sh) |
| OpenOil ensemble — 32 members | (cd scientific-workflows/openoil-ensemble-32 && ./prepare.sh && ./run.sh) |
| AI Assisted Analysis | (cd scientific-ai/ai-assisted-analysis && ./prepare.sh && ./run.sh) |
| Experiment Feedback Loop | (cd scientific-ai/experiment-feedback-loop && ./prepare.sh && ./run.sh) |
| Surrogate Model | (cd scientific-ai/surrogate-model && ./prepare.sh && ./run.sh) |
| Anomaly Detection | (cd machine-learning/anomaly-detection && ./prepare.sh && ./run.sh) |
| Batch Inference | (cd machine-learning/batch-inference && ./prepare.sh && ./run.sh) |
| Computer Vision | (cd machine-learning/computer-vision && ./prepare.sh && ./run.sh) |
| Hyperparameter Search | (cd machine-learning/hyperparameter-search && ./prepare.sh && ./run.sh) |
| Inference Benchmark | (cd machine-learning/inference-benchmark && ./prepare.sh && ./run.sh) |
| Infrastructure Selection | (cd machine-learning/infrastructure-selection && ./prepare.sh && ./run.sh) |
| LoRA Fine Tuning | (cd machine-learning/lora-fine-tuning && ./prepare.sh && ./run.sh) |
| Model Comparison | (cd machine-learning/model-comparison && ./prepare.sh && ./run.sh) |
| Model Conversion | (cd machine-learning/model-conversion && ./prepare.sh && ./run.sh) |
| Model Distillation | (cd machine-learning/model-distillation && ./prepare.sh && ./run.sh) |
| Model Ensemble | (cd machine-learning/model-ensemble && ./prepare.sh && ./run.sh) |
| Training with MLflow | (cd machine-learning/training-with-mlflow && ./prepare.sh && ./run.sh) |
| Audio Transcription | (cd generative-ai/audio-transcription && ./prepare.sh && ./run.sh) |
| Document Extraction | (cd generative-ai/document-extraction && ./prepare.sh && ./run.sh) |
| RAG Pipeline | (cd generative-ai/local-rag && ./prepare.sh && ./run.sh) |
| RAG Indexing | (cd generative-ai/rag-indexing && ./prepare.sh && ./run.sh) |
| Synthetic Dataset | (cd generative-ai/synthetic-dataset && ./prepare.sh && ./run.sh) |
| AI Video Studio | (cd generative-ai/video-studio && ./prepare.sh && ./run.sh) |
| Bounded Agent | (cd agentic-workflows/bounded-agent && ./prepare.sh && ./run.sh) |
| Multi-Agent Review | (cd agentic-workflows/multi-agent-review && ./prepare.sh && ./run.sh) |
| Supervised Tool Loop | (cd agentic-workflows/supervised-tool-loop && ./prepare.sh && ./run.sh) |
| Document Intelligence | (cd document-ai/document-intelligence && ./prepare.sh && ./run.sh) |
| Video Intelligence | (cd media/video-intelligence && ./prepare.sh && ./run.sh) |
| Blender Render Farm | (cd rendering/blender-render-farm && ./prepare.sh && ./run.sh) |
The standalone command writes outputs/manifest.json, outputs/dag-validation.json, and outputs/stage-<activity>.json. Independent activities become ready after their own dependencies complete; the local runner enforces a bounded number of concurrent processes. A failed stage prevents a successful final manifest.
Start a local AkôFlow daemon with Docker available and configure AKOFLOW_API_TOKEN in your shell. For example:
go run ./tools/daemon-local -example machine-learning/training-with-mlflowThis validates Docker through the daemon API, builds the declared OCI image, registers local-only infrastructure and the workflow, imports a plan, and submits a real run. The daemon performs execution and artifact transfers; the renderer never runs connectivity checks, shell commands, or Docker. The command records the exact projection in evidence/daemon-run.json and prints the run ID for inspection in Desktop.
For the full long-running scientific example, submit and leave the daemon in charge:
go run ./tools/daemon-local -example scientific-workflows/montage-050 -wait=falseThe local profiles are explicit about their inputs and algorithms. They include numerical integration and regression, a tiny rank-one adapter, geometric image classification, real ONNX and TorchScript export, schema-based generation, a tone-vocabulary audio recognizer, pixel-template OCR, and bounded deterministic tool agents. Infrastructure comparisons separate CPU measurements from declared placement assumptions. The MLflow example uses the actual SDK and a local SQLite tracking store.
Montage preserves all 58 original native commands and 138 named file handoffs from the pinned 050d image. It is not a timing simulation and can take substantially longer than the small profiles. Its CI preparation check verifies the source and image contract without claiming full mosaic completion.
scientific-workflows/andscientific-ai/: astronomical mosaics, oil-spill ensembles, numerical experiments, and learned scientific surrogates.machine-learning/,generative-ai/,document-ai/,media/,rendering/, andagentic-workflows/: local domain examples.workloads/: canonical implementations for the 22 small pattern workloads;go run ./tools/materializegenerates their standalone source bundles.tools/: Go runners, catalog discovery, infrastructure generation, documentation export, and validators.simulation/,local/,kind/,slurm/,real/,connections/,onboarding/,server-instance/, andexperiments/: additional infrastructure and experiment bundles.
The showcase standard defines offline workload execution, real computation, observable activities, and artifact validation. Product documentation is published at akoflow.com.
The example files live in this repository, while the reusable CI implementation
is maintained in UFFeScience/akoflow/.github/workflows/examples-ci.yml. The
small workflow in this repository passes the exact pull-request or main
commit to that central workflow. Runnable showcases are discovered from their
docker/Dockerfile and executable run.sh; adding a showcase does not require
editing a duplicated matrix.
Automatic PR and push runs select only showcases whose own directories changed. A PR compares against its merge base; a push compares against the previous commit of that push, including every commit in a multi-commit push. Changes outside the showcase directories run the shared tooling checks without executing workloads or rebuilding images. Renames include both affected directories; removed showcases are skipped. Superseded runs of the same PR are cancelled.
Use Actions → Example checks → Run workflow → all_examples for an explicit
full-catalog run. Without that option, a manual run selects changes in the latest
commit. The reusable workflow also accepts examples_base_ref to select another
comparison range.
After a push to main, CI publishes only the selected validated showcase images to GHCR and
notifies the AkôFlow documentation workflow. Configure the repository secret
AKOFLOW_DOCS_DISPATCH_TOKEN with access to dispatch workflows in
UFFeScience/akoflow. If the secret is temporarily unavailable, the daily
documentation reconciliation still checks out this repository and validates
all referenced example files.
These seven examples adapt studies from the Brazilian e-Science Workshop (BreSci), 2024–2026, co-located with SBBD. Each README credits the original authors, links the paper and original data/code, describes the exact reproduced scope and protocol adaptations, and embeds actual result figures plus an observed native execution timeline. Input SHA-256 receipts, completed daemon evidence and independent artifact validators accompany each bundle.
Start from any linked bundle and run ./prepare.sh && ./run.sh. Preparation acquires original inputs; computation executes offline. Native daemon execution instructions are in each README. Complete original inputs are used within each explicitly selected scope; the MAFAULDA portable profile and TPC-H Q1 audit do not claim whole-paper reproduction. Measured overlap demonstrates actual parallel execution, without assuming a speedup over a sequential baseline.
The Estrada Real search stage reproduces all 3,715 reported Greedy divergences; Eidolon's three rebuilt full R collections match the pinned published references. The database example preserves and exposes the different Q1 date cutoffs in the original code before testing an explicitly aligned variant.
BreSci Hub discovery tags: bresci, bresci-2024, bresci-2025, and bresci-2026. Each publication retains its original conference/year and scientific author attribution.