Frontier AI forthe applied sciences

Backbond is a product by Optagon Labs, a research and product company building frontier AI for applied sciences.

Research spans inference-time compute, optimization under data scarcity, verification, scientific knowledge representation, transfer learning and multi-agent systems.

Optagon Labs

Optagon Labs is a research and product company. Research questions originate in formulation programmes that are running, and results return to Backbond on the same cycle.

The six areas below are the constraints that recur across those programmes. None of them is resolved by model scale alone, and each is stated here at the level of the open problem rather than the implemented feature.

01Inference-time compute

Symbol: Elected path
Symbol: Candidate branches
Symbol: Compute spent

Figure 01. Compute allocated across a tree of dependent inferences. The accent path is the subproblem the policy elects to search further; the bar beneath is cumulative compute, which is spent unevenly by design.

Inference-time compute for scientific decision making

A formulation decision decomposes into long chains of dependent inferences, most of them inconsequential and a few of them determining the outcome. We study the allocation of inference-time computation across such chains: identifying high-leverage subproblems, scheduling additional search against them, and setting termination criteria on decision value rather than token budget. The same work covers loop embodiment, where the policy selects which simulations to run and which experiments to propose rather than answering from its priors, and where intermediate steps are retained so a conclusion can be audited by the scientist accountable for it.

02Low-data search

Symbol: Posterior mean
Symbol: Credible interval
Symbol: Observation
Symbol: Proposed experiment

Figure 02. Posterior mean and interval over one design parameter, conditioned on four observations. The proposed experiment is sited where expected information gain is highest, which is not where the mean is best.

Optimization and search in low-data regimes

Formulation programmes typically hold 10² to 10³ measurements against design spaces of 10⁹ or more, and each additional observation costs weeks of laboratory time. We develop optimizers for that regime: physics-constrained surrogates that stay well-posed outside the sampled region, acquisition functions selected for expected information gain rather than expected improvement, and uncertainty quantification calibrated tightly enough that reported intervals hold under held-out validation. Batch and multi-objective settings are the default, since experiments are run in parallel and no single objective is decisive.

03Verification loops

Symbol: Generator, verifier
Symbol: Cleared assertion
Symbol: Returned with constraint

Figure 03. Generation separated from verification. A claim either clears the gate or returns carrying the constraint it violated, so the next attempt is conditioned on the failure rather than resampled from the same distribution.

Verification loops

Generation and verification are separated. Candidate claims are decomposed into checkable assertions and tested against mechanistic constraints, simulation output and the literature, with adversarial passes that attempt refutation before an assertion is admitted. Failures return the violated constraint, so the next attempt is conditioned on the failure rather than resampled. Open questions include which classes of scientific claim admit automatic verification at all, how to bound residual error on the classes that do not, and how to keep verification cost sublinear in generation volume.

04Knowledge representation

Symbol: Ingredient
Symbol: Unit operation
Symbol: Measured property
Symbol: Traced inference

Figure 04. A typed heterogeneous graph over ingredients, unit operations and measured properties. The accent path is one inference walked across it; every edge carries provenance and evidence strength, so a conclusion inherits the quality of what it rests on.

Knowledge representation and chemical digital twins

Composition, unit operations, structure and measured performance stand in typed relationships, and those relationships carry the mechanism. We study representations that make them explicit and computable: heterogeneous graphs over ingredients, processes and properties, with provenance and evidence strength attached to every edge so a downstream inference can be weighted by the quality of what it rests on. On that substrate we build chemical digital twins that resolve a formulation at molecular, interfacial, morphological and environmental scales simultaneously, since the scales are coupled and a change at one propagates to the rest.

05Transfer learning

Symbol: Source domain
Symbol: Target domain
Symbol: Transferred structure

Figure 05. Shared structure fit where observations are dense, carried to a domain holding few. Establishing where the analogy fails is the harder half: a transferred prior either helps or silently biases the search.

Transfer learning

The physics governing an emulsion, a porous scaffold or a barrier coating is invariant to the product built on it. We study transfer at the level of mechanism rather than correlation: shared latent structure across products, programmes and industries, adaptation to target domains holding a handful of measurements, and detection of negative transfer. The last is the harder half, because establishing the conditions under which two systems are not analogous determines whether a transferred prior helps or silently biases the search.

06Multi-agent harnesses

Symbol: Specialist agent
Symbol: Orchestrator
Symbol: Typed hand-off

Figure 06. Specialised agents under deterministic orchestration. Hand-offs are typed and logged, and disagreement between agents is retained rather than averaged away, since the disagreement is usually the finding.

Multi-agent architectures and harnesses

Formulation work decomposes into literature synthesis, simulation, property prediction, manufacturability assessment and experimental design. We study architectures that decompose it the same way, and the harnesses that make a fleet of specialised agents tractable: deterministic orchestration of parallel and blocking stages, typed structured outputs at every hand-off, disagreement between agents retained rather than averaged away, and audit trails from a result back to the agent and evidence that produced it.

For research collaborations, data partnerships or open roles, write to research@optagonlabs.com.

Open roles