AI systems
One architect, just over a year, 20+ AI solutions: an enterprise-shaped AI platform, designed end to end and running on a private lab I built myself. It has what an enterprise AI program needs: clear layers, one governed front door for every model, and open standards between the parts (MCP, OpenAI-compatible APIs and A2A, the emerging agent-to-agent protocol), so no vendor owns it. Every component states its real status. It is the kind of platform I would design for your organization, and I have already built it.
Architecture
The separation of concerns an enterprise AI platform needs to stay governable as it grows: governance at the edge, agents in the middle, memory and knowledge underneath, and business applications on top.
Design principles
Each rule answers a question every Head of AI eventually gets from the business: can we audit it, what does it cost, who owns our data, and are we locked in? They run through the design of every component, and they are the rules I would bring to your AI program.
Deterministic where it matters
Answers an auditor can reproduce. Evidence and memory, the layers other systems rely on, are deterministic: the same input gives the same answer, with no language model in the loop. AI is used where judgment helps, not where an audit needs a fixed answer.
Evidence before claims
Traceability by default. Every repository carries a machine-readable evidence pack describing what it is and what it does, and knowledge is published only after it passes an evidence gate.
Governed gateway
Cost control that finance will sign off on. All model traffic goes through one front door, and every call leaves a spend receipt, so AI cost is visible and controllable the way any other operating cost is.
Local-first, vendor-independent
Data privacy and vendor independence. The ecosystem runs on local models in my own lab and can route to cloud APIs when that is the better choice. Swapping a model or provider is a configuration change, not a rewrite.
Open standards
No lock-in, at any layer. Components talk through MCP, OpenAI-compatible APIs and A2A, so they work with the major AI coding agents and with each other, and none is tied to one vendor.
Honest status labels
Decisions made on facts, not demos. Each component states its maturity plainly: active development, pre-release, research or design phase. Research is labelled as research, and negative results are kept, not hidden.
Flagships
Evidence, memory, optimization and the gateway: the four capabilities an enterprise AI platform cannot do without, and the pieces every other component here depends on.
Flagship · Evidence
AI agents that work from facts, not guesses. DaVinci is a deterministic engine that turns any codebase into an evidence pack agents can trust, with no LLM guessing. It powers the Wiki: 33,000+ open-source repositories and 1B+ lines of code analyzed.
Flagship · Memory
Agent memory a compliance team can stand behind. Deterministic, compliance-grade memory for AI agents, with no LLM in the loop, in one binary that plugs into the major coding agents (Claude, Codex, Gemini and more) over MCP.
Flagship · Optimization
Hard optimization problems, solved without specialized hardware. An optimization engine (QUBO/Ising) that runs on consumer GPUs and decides what context an AI model sees: a supply-chain-style selection problem, applied to AI.
Flagship · Gateway
One governed front door for every model, with no silent spend. An OpenAI-compatible gateway that routes work across local models, cloud APIs and agent tools, across six execution surfaces, and leaves a spend receipt for every call.
The wider portfolio
A portfolio managed the way an enterprise should manage its AI investments: older experiments and duplicates retired or archived, and every component still in play carrying an honest status label.
Agent orchestration
Agents you can let loose without losing control. An agent operating system that treats every model as untrusted: agents run behind a security boundary, with verifiable identity and an audit trail, from a command line or a desktop app.
Pre-release
Agent orchestration
Designed so agents your teams already have can work together instead of being rebuilt. The shared protocol layer: common contracts so agents and tools built on different frameworks (such as LangChain and AutoGen) can work together, including A2A agent cards.
Design phase
Agent orchestration
A platform that assembles itself, even offline. The connective tissue: a registry that lets each component find and call the others, and can set up the platform without an internet connection.
Active development
Agent orchestration
AI-written code that passes controls before it lands. An AI coding environment with governance built in: proposed changes pass a veto check and signed manifests, and DaVinci evidence is wired in as context for the agent.
Active development
Memory
The proving ground behind AgentMemory. LLM-assisted long-term memory for agents (facts and mental models), built on the open-source Hindsight engine. The testbed whose comparisons shaped AgentMemory's design.
Research
Knowledge
Vetted knowledge on tap for every agent. The knowledge base: 33,000+ open-source repositories catalogued from DaVinci evidence (1B+ lines of code analyzed), published only after an evidence gate, refreshed daily and served to agents over MCP.
Active development
Optimization research
Aimed at lowering the hardware bill for running AI. Research on using the optimization engine to prune large language models so they run on smaller hardware. Experiments are pre-registered and negative results are documented.
Research
Model training
Turning the platform's own data into a specialist model. The model-training lane: data contracts that bind memory, evidence and optimization into one pipeline for training a specialist coding model.
Research
Gateway & governance
Designed to keep sensitive data from leaking through AI. A zero-trust security layer for AI traffic that screens prompts and outbound data. It is being consolidated to sit behind Connect as its enforcement layer.
Design phase
Industry vertical
AI for regulated engineering work, where every number must be defensible. Verifiable engineering calculations: 182 calculation blocks across 17 engineering disciplines, with every result written to a tamper-evident audit ledger and graded on an evidence ladder.
Pre-release
Industry vertical
Forward-looking intelligence, with risk controls designed in from the start. Research into prediction-market intelligence, run paper-only with no real money, with a kill switch, operator approval and circuit breakers in the design. Its timestamping component, Attest, is published as a Python package.
Research
Industry vertical
Designed for agents doing production work, with an explicit sign-off before release. An agent-driven website production line, with adversarial checks on the output and an explicit authorization step before anything ships. It runs its model calls through Connect.
Design phase
Where it runs
Most architects rent their AI. The whole ecosystem runs on a six-machine lab I built myself: local models, agent fleets and evaluations around the clock, with no data leaving my control. The same discipline an enterprise needs for data privacy, cost control and vendor independence.
Don't take my word for it. Paste the role description and get an honest read on whether I fit: evidence, gaps and a recommendation. That includes "not a fit" when it's true.
