The intelligence layer
for modern AI training.
Metrana captures high-dimensional training signals and turns them into actionable insight. Built for complex systems: from multi-agent reinforcement learning to large-scale LLM training.
From signal to understanding — across your entire training system.
Capture everything. At any scale. At any resolution.
Logging that captures what other tools force you to drop. Modern training runs produce thousands of metrics across layers, environments, and agents. Metrana ingests and structures them without forcing you to sample, compress, or drop information — and does it without exponential cost growth as workloads scale.
From dashboards to diagnosis.
Metrana doesn't just visualise your system, it diagnoses it. An embedded agentic layer continuously analyses training dynamics to surface issues and root cause in real time — so you're not left reading charts and guessing.
Built for the complexity of modern AI systems.
- Multi-agent reinforcement learning - Track per-environment signals, rewards, policies, and trajectories. Understand emergent behaviour and debug instability across agents.
- LLM Training - Monitor loss, gradients, activations, and scaling behaviour. Diagnose instability in large-scale distributed runs.
- General Machine Learning - Works across any training pipeline. Integrates with PyTorch, TensorFlow, and JAX. Supports research and production environments.
Integrate in minutes, not days
import metrana
metrana.init(workspace_name="gpt2-master", project_name="gpt2-training", run_name="gpt2-dist-lr0.00001")
metrana.log("train/loss", loss)
metrana.close()Metrana fits directly into your existing training pipeline with minimal code — comparable to Weights & Biases.
metrana.init(project_name=PROJECT, workspace_name=WORKSPACE, run_name=RUN_NAME)
metrana.log_rl_step("train/reward", reward)
metrana.log_rl_episode("train/final_cumulative_reward", reward, rl_step=rl_step, env_id=env_id)
metrana.log_rl_environment_step("train/reward", reward, episode=episode, rl_step=rl_step, env_id=env_id)
metrana.close()Metrana also makes it easy to log signals at different timescales within RL systems — from individual steps to full episodes and across parallel environments.
Works with your existing stack
Metrana is designed to integrate cleanly into real-world systems
- drop into existing training loops without refactoring
- compatible with common ML frameworks and custom pipelines
- works alongside your current tooling — no need to replace your stack
Flexible integration paths
Whether you’re running simple experiments or complex orchestration, Metrana supports:
- lightweight SDK-based integration for rapid setup
- deeper instrumentation for full system coverage
- incremental adoption — start small, expand as needed
AI-Assisted integration and metric discovery
Metrana’s agentic assistant handles both integration and metric selection — particularly valuable for advanced training workflows:. Its integration agent can -
- add the required logging code directly into your codebase
- suggest a comprehensive set of metrics to track based on your training setup
- identify missing signals that are critical for understanding system behaviouR
Four steps from raw signal to confident action.
Ingest
Log high-volume metrics directly from your training runs in real time.
Structure
Organise data across models, environments, and components into a unified system view.
Analyse
Interpret system training dynamics, detect anomalies, and identify emerging failure modes.
Recommend
Surface insights, explain behaviour, and suggest targeted optimisation actions.
Built for the scale and complexity of modern AI training
System-level visibility
Operate complex training systems with full visibility. Metrana structures thousands of signals into a coherent system view so nothing gets lost between components.
Built for multi-agent complexity
Track per-environment signals, rewards, and trajectories across every agent in your system. When behaviour emerges or breaks, you see exactly where and why.
Faster diagnosis and resolutions
Fix problems faster with clear, actionable recommendations. Metrana traces failures to their origin, not the symptom, so you know what started it and when.
Decisions grounded in data
Pinpoint root causes and take decisive action. Every recommendation comes from what the system is actually doing. Not heuristics, not guesswork.
Backed by leading deep-tech investors
