01
Prompt
Start with an idea you want to explore.
Faster training for vertical frontier models in physics · engineering · finance
Accelerate your research.
Agents survey your IP and the literature, propose training-run directions, and log every benchmarked result to an immutable graph that future runs can build on.
Plan
Run
Evaluate
Publish
Trusted by ETH Zurich
The problem
Iteration speed is bottlenecked by the ability to try new ideas informed by past research.
Generic experiment trackers log metrics. Methodic captures the entire decision graph behind every run and makes it legible to agents and researchers — informing experiment planning and evaluation.
A defensible training-data flywheel for an AI-native world.
Metric ● Status quo ● Methodic
Iteration speed
Agents propose, launch, and benchmark training runs end to end.
Institutional memory
Every experiment's design, lineage, and results are recorded. Regressions are investigated and documented to inform future runs.
Researcher onboarding
New researchers leverage accumulated experimental knowledge from day one.
How it works
Go from idea to shared knowledge.
02
Explore
Agents survey internal IP and external literature, then propose promising training-run directions.
03
Benchmark
Launch and log training runs against your eval suite. Conditions, metrics, and artifacts captured.
04
Publish
Commit results to your immutable experiment graph. Future runs build on past runs.
Experiment graph
Every experiment becomes a node that informs future directions.
→ Immutable
Once an experiment is committed, its design is frozen. Once it completes, its outputs (inputs, hyperparameters, metrics, and artifacts) are locked.
→ Graph lineage
Each run records the prior runs it descends from. Improvements over a baseline are explicit, not narrated after the fact.
→ Searchable
Find every prior run that touched the same dataset, eval, or architecture — across teams, across years.
→ Auditable
Connect datasets to every downstream trained model to streamline compliance reporting.
Run lineage · vert-fm-radiology
r-08c1
Pretrained checkpoint, no domain adaptation
eval/radiology-bench: 0.612 · parents: ∅
r-12a4
Finetune on internal report corpus (v3)
eval/radiology-bench: 0.741 (+0.129) · referenced by 6 downstream runs
r-19f7
Aligned to radiologist preference pairs
eval/radiology-bench: 0.689 (−0.052) — regression vs parent.
r-1bd2
Aligned with rebalanced preference set
eval/radiology-bench: 0.778 (+0.037) · promoted to candidate-prod.
Example workflow — illustrative until a real pilot is published.
Security & compliance
Built for regulated environments.
ACCESS
Role-based access control
Org-wide RBAC, IdP-mapped granular roles.
ENCRYPTION
Encrypted at rest and in transit
AES-256 / TLS 1.3, per-tenant keys.
AUDIT
Fully audited actions
Append-only audit log of every action and configuration change.
COMPLIANCE
Lineage built for regulators
Immutable graph streamlines regulatory report generation.
FAQ
Frequently asked.
Generic LLM tools optimize for fluent text. Methodic optimizes for training-run iteration on vertical-specific frontier models — every run, dataset, eval, and artifact lives in one immutable graph future runs can build on.
Accelerate your model development.
Create your account. We're onboarding 1–2 research teams per week.