Dawn State
an applied AI studio
Dawn State builds domain-specific models for complex physical and economic systems.
focus areas
Project STROBE — Protein Dynamics & Cryptic Pocket Engine
Static structural models (e.g. AlphaFold) capture rigid ground states but miss the dynamic motions—loop shifts, helix displacements—that reveal transient cryptic binding pockets. Traditional Molecular Dynamics (MD) can capture these motions, but microsecond trajectories are computationally prohibitive at proteome scale.
STROBE (Structural Thermodynamic Representation & Open Pocket Engine) uses high-throughput generative AI thermodynamic sampling to "flash" the protein's thermodynamic landscape. It outputs 3D PDB coordinates of open states ready for virtual screening and binder design in plant metabolism, crop protection, and enzyme engineering.
Technical Architecture & Capabilities
3D Coordinate Output: Unlike SOTA classifiers that only predict where a pocket might form, STROBE generates physical 3D PDB coordinates of the open state.
Ligand-Agnostic Discovery: Samples intrinsic apo-side flexibility without requiring a known cofolding ligand, uncovering uncharacterized allosteric and cryptic sites.
Proven Calibration: Verified on 39 experimentally validated apo/holo pairs with 27 judge-verified open-state captures, scaling across proteome-scale agricultural and fungal targets.
Multi-Agent Systems — Game Theory & Population Dynamics
Markets, negotiations, and strategic decisions are complex adaptive systems driven by distributed incentives, asymmetric information, and dynamic agent interactions. We build multi-agent simulation frameworks to study equilibrium dynamics, adversarial alignment, and evolutionary self-improvement in neural policies.
Our multi-agent research spans two core pillars: Adversarial Behavioral Auditing under economic duress, and Continuous Learning Through Weight-Space Model Recombination.
Adversarial Alignment & Synthetic Data Loops
Alignment Under Duress: Standard frontier LLMs exhibit situational alignment—when subjected to game-theoretic negotiations with asymmetric information, their safety guardrails dissolve under financial pressure.
Adversarial Harvesting: We use a proprietary framework to design game-theoretic environments that harvest high-variance tactical deviations and produce high-quality synthetic training data.
Mech-Interp & Anti-Fragility: By feeding synthetic failure modes into mechanistic interpretability pipelines, we analyze internal steering activations while models operate under duress, baking behavioral boundaries directly into model weights rather than relying on fragile system prompts.
Continuous Learning Through Weight-Space Model Recombination
Beyond standard gradient descent, we explore weight-space recombination across neural populations. By combining trained model sub-structures across specialized environments, agents can assemble complementary capabilities in a single generation without re-training from scratch.
get in touch
email matt@dawnstate.xyz
github/dawn-state github/matty-west shelltoho.me linkedin