Institutional memory
every tried condition stays queryable, failures included. Negative results stop being lost.
Science & R&D
Industry applicationA good R&D assistant isn't a smarter model — it's a model that finally reaches the link between what you set out to test (hypotheses, Design of Experiments) and what actually happened (results). That link is the substrate. Without it, lab AI is a literature chatbot. With it, every experiment makes the next one faster.
The problem
Experimental designs live in spreadsheets. Results live in instrument exports, electronic lab notebooks, scattered files. And the reasoning — why this condition, what the result meant — lives in people's heads, and leaves with them. Negative results vanish. "Didn't we already try that?" has no answer.
Put an AI assistant on top of this disconnection and it inherits it: it can summarize a paper, but it cannot reason about your program, because your program was never captured as connected knowledge.
Lab AI that gets it wrong doesn't have a model problem. It has a foundation problem.
The scientific substrate
Syrtis captures the experimental loop as connected, sovereign knowledge: the hypothesis, the design (factors, levels, conditions — the Design of Experiments), the execution, the measured result, and the scientist's interpretation. Each design point is bound to its outcome and its rationale. This connected corpus is three things at once:
every tried condition stays queryable, failures included. Negative results stop being lost.
the R&D assistant reasons over what you actually did, cited and traceable, not generic web knowledge.
enough (design → result) pairs let you model the response surface, propose the next experiment, and close the loop (active learning, Bayesian optimization).
"A result alone is data. A result tied to the intent that produced it is knowledge."
Every designed experiment is bound to its result and its rationale. The substrate keeps the loop alive — and the assistant proposes the next iteration.
The R&D assistant
Sitting on the substrate, the assistant reads the exact context it needs and writes fresh knowledge back — the same pattern as Syrtis agents, applied to the bench.
Suggests the next conditions from the response surface — not guesswork.
Flags conditions already explored, failures included, with the run that proves it.
Drafts the report or protocol, every claim cited to its run.
Surfaces inconsistent results across batches, operators, instruments.
Onboards a new scientist into years of program memory in days.
Where it applies
The pattern is domain-agnostic: wherever a team designs experiments and reads measured outcomes, the substrate applies. For example —
process and recipe optimization (etch, deposition, lithography), yield ramp, DoE across tool parameters.
excipient screening, co-formulation, stability.
alloys, composites, coatings: composition × process × property.
reaction optimization, ligand and condition screening.
electrolyte/electrode formulation, cycling outcomes.
fermentation, scale-up, media optimization.
Different molecules, the same loop — hypotheses designed, results measured, knowledge that should compound but usually evaporates.
Sovereignty & compliance
R&D data is the crown jewel: intellectual property, trade secrets, partner confidentiality. Syrtis runs on-premises or in sovereign EU cloud. Model-by-node selection keeps sensitive steps on local models. Full provenance, versioning, audit trail. GDPR and the EU AI Act respected by design. This is also what makes deployment possible behind confidentiality walls and inside consortium R&D.
Through ingestion nodes: documents, SQL databases, APIs/webhooks, instrument exports. Everything enters via the indexer, with no upfront rebuild.
No. Capture comes first; indexing routes each item into the memory that fits it (vector, graph, document, SQL).
No. Syrtis links and remembers the design → result pair. It complements your DoE and optimization, it doesn't replace them.
In a sovereign substrate: on-premises or EU cloud. Data does not leave your perimeter.
A first assistant is built in hours, not months. Compounding value grows as the experimental loop fills in.