MLos DomainOS¶
Status: Production
Engine: aiDevops
Domain: Deep learning for genomics (Karpathy 6-phase methodology)
GPU-aware deep-learning DomainOS for genomics. Structures every training run with Andrej Karpathy's six-phase recipe and auto-detects framework and accelerator (NVIDIA CUDA, Apple MPS, Google TPU) with Weights & Biases integration.
| Field | Value |
|---|---|
| Product | MLos |
| Engine id | aiDevops |
| Intent key | model_training |
| Repository | Hordago-Labs/aiDevops |
| Plugin | ml-os |
| Maturity (registry) | Production |
Overview¶
GPU-aware deep-learning DomainOS for genomics. Structures every training run with Andrej Karpathy's six-phase recipe and auto-detects framework and accelerator (NVIDIA CUDA, Apple MPS, Google TPU) with Weights & Biases integration.
Scientific approach¶
MLos structures every run as 6 phases. Gates marked HITL require an explicit human-in-the-loop approval before the run advances.
| # | Phase | Gate | HITL |
|---|---|---|---|
| 1 | Data Inspection | Inspect and understand the dataset. | No |
| 2 | End-to-End Skeleton | Wire a minimal training/eval loop. | No |
| 3 | Overfit Verification | Prove the model can overfit a small batch. | No |
| 4 | Regularization | Introduce regularization and augmentation. | No |
| 5 | Hyperparameter Tuning | Sweep hyperparameters. | No |
| 6 | Final Optimization | Squeeze final performance and freeze. | Yes |
Capabilities & evidence objects¶
Domain tools / skills
jax-trainingpytorch-trainingtensorflow-trainingkarpathy-recipevariant-effect-predictionprotein-embeddinggpu-setupwandb-integration
Evidence objects
| Object | Role | Consumer |
|---|---|---|
result.json |
produces | co-writer |
provenance.json |
produces | evidence-audit (framework + GPU + seed lineage) |
Canonical artifacts (5-artifact contract)
Emitted
MLos emits the full 5-artifact contract on every completed run.
| Artifact | Description |
|---|---|
result.json |
Structured primary result payload for the domain run. |
report.md |
Human-readable narrative summary of the analysis. |
provenance.json |
Tool versions, reference data, and algorithm lineage for reproducibility. |
gate_status.json |
Per-phase gate pass/fail decisions. |
session_summary.json |
Session metadata for replay and audit. |
Standalone quickstart¶
Zero platform dependency
This quickstart runs the DomainOS standalone. The Hordago platform is not required; platform composition is opt-in (see Composition below).
- Install the standalone
ml-osplugin (Hordago-Labs/aiDevops) -- no Hordago platform required. - Invoke the domain skill with a MLos intent (see the intent keywords below).
- Review the emitted artifacts under the run's output directory.
Intent keywords (route to this engine):
train fine-tune model deep learning neural network embeddings variant effect predict
Worked example¶
Train a variant-effect-prediction model on dual RTX 5090s
A user asks to train a variant-effect predictor. MLos runs detect_gpu.py, selects a framework, then walks the six Karpathy phases from data inspection through final optimization, logging to W&B and emitting training artifacts.
Validation & benchmarks¶
Benchmarks
- Overfit-a-batch sanity gate
- Held-out AUROC regression
Reproducibility. provenance.json pins framework versions, random seeds, and GPU topology; runs are replayable from the artifact bundle.
Reference¶
MCP fallback servers (used when the plugin is unavailable): pubmed
Source documents
- MLos integration:
references/mlos-integration.md - Engine catalog (Tier 2b):
references/engine-catalog.md
Composition¶
Platform opt-in
MLos runs standalone. When composed under the Hordago platform it gains cross-domain routing, the Shared Compiler gate, and evidence-audit provenance enforcement. Platform composition is opt-in; the quickstart above has zero platform dependency.
Cross-domain dependencies
| Engine | Relationship |
|---|---|
struct-os |
Protein-embedding models feed structure-prediction workflows. |
onco-os |
Variant-effect-prediction models score somatic variants. |