Building the stack: TriMemory resolves what the agent knows, veronica-core enforces that the inference stays within budget.
TriMemory -- Memory architecture for LLM agents
- Three-path memory: KV window + retrieval index + TRN recurrent state
- 8 KB per agent at any context length (vs 156 MB KV cache at 10K tokens)
- 1,000 concurrent agent states in 16 MB
- 277 tests, Apache 2.0
VERONICA Core v3.7.6 -- Runtime containment for LLM agents
- Hard budget, step limits, retry caps, circuit breakers -- evaluated before the call reaches the model
- Memory governance, message hooks, DEGRADE directives, semantic loop detection
- Contributor to AG2 v0.11.3 (PR #2430), LangChain / CrewAI / LangGraph / LlamaIndex adapters
- 6125 tests, 94% coverage, zero required dependencies
pip install veronica-core
VERONICA v0.8.1 -- LLM governance control plane
- Policy authoring, simulation, rollout pipelines
- Tenant hierarchy, incident replay, audit dashboards
- Built on veronica-core
- 1197 tests
pip install veronica-cp
Construction consultant, 15 years. Regulated environments where failure has real cost. Now applying that mindset to LLM agent safety.
Python, C++, CUDA, TypeScript, WebGPU
Runtime containment, multi-agent orchestration, GPU kernels, 3D Gaussian Splatting
- The $0.64 bug: how nested retries silently multiply your LLM costs (dev.to)
- Zenn articles -- CUDA bugs, 3DGS, LLM agent cost control (Japanese)
Available for remote contract work -- AI safety, agent infrastructure, runtime systems.