Research

Research that becomes infrastructure.

Thyn Research studies the systems problems behind local intelligence, autonomous software, market engines, and verified execution.

Research agenda.

Area 01
Local Intelligence Systems

How much cognition can run near the user?

We study the boundary between local inference, cloud-scale reasoning, private context, and hybrid orchestration.

  • On-device model routing and scheduling.
  • Private context windows and memory stores.
  • Latency-aware model cascades.
Area 02
Agent Verification

How do autonomous systems prove they did the right thing?

Reliable agents need more than tool calls. They need traces, policies, invariants, testable plans, and failure recovery.

  • Plan verification and rollback semantics.
  • Regression harnesses for agent behavior.
  • Policy-constrained execution.
Area 03
Market Intelligence

How should AI reason inside adversarial markets?

Trading systems require simulation, timing, uncertainty estimation, risk controls, and adversarial robustness.

  • Execution strategy simulation.
  • Market microstructure-aware evaluation.
  • Risk-bounded autonomous actions.
Area 04
Cryptographic Execution

What should software be able to prove?

Cryptographic infrastructure offers a path toward stronger guarantees around identity, history, replay, signatures, and authorization.

  • Verifiable logs and replay systems.
  • Policy proofs for financial workflows.
  • Self-hosted trust boundaries.
Area 05
Growth Systems

Can marketing become an experimental control system?

Autonomous growth systems combine creative generation, attribution, experiment design, budget allocation, and policy constraints.

  • Causal attribution loops.
  • Content generation with brand constraints.
  • Campaign simulation and optimization.

Research style.

We prefer testable questions over category language. A good Thyn research question should have a workload, a metric, a failure mode, and a path to production.

A useful research program has:

1. measurable workload
2. explicit constraint
3. reproducible baseline
4. product path
5. failure analysis

Otherwise it is just a narrative.

Research notes.

Note

Latency-aware intelligence

Why AI products need latency budgets, not only model benchmarks. A model that scores well offline can still feel broken when the full loop misses the time the user will actually wait.

Note

Replayable agents

A framework for traces, deterministic tests, and failure analysis in autonomous workflows. When every action is replayable, a bad run becomes a reproducible case instead of a one-time mystery.

Note

Market simulation loops

How execution engines can evaluate strategies before entering live environments. Simulating against historical and synthetic order flow surfaces risk and slippage while mistakes still cost nothing.

Bring intelligence closer.

See open rolesRead the technology