Systems that think clearly.
A boutique consulting practice, a product lab, and a small AI research lab. We build production AI systems for operators — voice agents that answer the phones, multi-agent systems that run the back office, and development workflows that get engineering teams running on tools that actually keep up.
One operator. A small number of slots each quarter. The first conversation is a thirty-minute call where we tell you in plain English whether AI will help — or whether you are being sold a fantasy.
What we build for operators.
Agentic voice workflows.
Voice is the door. The agents do the work. A multi-agent workflow with voice in front — the caller is qualified, then handed off in real time to the specialist agent or the human who can resolve the call.
Built on ElevenLabs voice infrastructure.
The operations system.
Back-office work — email triage, scheduling, follow-up, weekly reporting — run by a multi-agent system that knows your business and asks before it ships anything irreversible.
The development workflow audit.
Your engineers are using AI coding tools, but the workflow around them is leaking velocity. We audit, redesign, build the custom skills and CI gates, and leave the team running on a workflow that keeps up.
What we work on underneath.
Underneath the practice sits a small AI research lab. The work here is what keeps the practice and the product lab honest.
Local inference — running frontier-class models on hardware the customer owns. Privacy as architecture, not policy.
Fine-tuning — adapting general models for narrow, mission-critical work. A precise tool, not a default.
Security — adversarial review, evidence-bound CI, supply-chain audits for agent systems.
Memory systems — semantic memory, knowledge graphs, long-context retrieval. Memory you own.
Agent harness — the scaffolding around agents that makes them reliable. Tool tiers, mechanical enforcement, eval loops.
Evals — designing the tests that tell us whether an agent actually works. Falsification protocols, null model controls, hold-out prediction. The verification side of the harness.
Robotics — agentic loops applied to physical systems. Early.
Agentic voice — voice as the primary surface for agent work, not a chat-window add-on.
What we are building of our own.
Vignos.
AlphaFoundational agent. Local inference, memory, and security in one package. The substrate the practice will run on as it matures.
Bandit.
Pre-releaseRepo-native trust layer for agentic software delivery. Sits between the agent and the code, with mechanical enforcement at every gate.
Seen Together.
BetaCouples counselling AI for partners navigating hard conversations, with a therapist in the loop.
Bimblepath.
LiveMac planning for variable-energy days. Outcomes, capacity-aware Today planning, and commitments on the horizon. One-time purchase, no subscription.
Bring a hard problem. Leave with a system.
The first conversation is a thirty-minute call. We tell you in plain English whether AI will help your business — or whether you're being sold a fantasy.