OneNomad Studio / AI practice

The studio builds AI into real products and teaches teams to work with it. No slideware, no magic claims. Working software, honest limits, and skills your people keep.

What the studio does with AI

Real AI capability shipped into products, plus the training to run it without a vendor on retainer.

FEATURES

AI product features

Retrieval and RAG, vision (receipt scanning shipped in production with Gemini Vision), generation, and chat interfaces. Built into an existing product or a new one.

RAGVisionGenerationChat

AGENTS

Agent systems

Multi-agent workflows with budgets, test gates, approval steps, and kill switches. The studio runs its own agent crew to build software, so the failure modes are known first-hand.

OrchestrationGuardrailsEvalsMCP

KNOWLEDGE

Company knowledge

Connect internal tools and docs to AI with permissions enforced at the data layer, including on-prem when the data cannot leave.

MCP serversRetrievalAccess controlOn-prem

TRAINING

Team enablement

Hands-on training for engineering and product teams: AI coding tools, agent workflows, writing briefs and review gates that keep quality up. Skills stay when the engagement ends.

WorkshopsPairingPlaybooksAdoption

Not demos. Shipped.

Every item below is real software people use, built by this studio.

przm Memory

Live

Long-term memory for AI agents: hybrid retrieval across vectors, BM25, and a typed knowledge graph. Scores 92% R@10 on the LoCoMo benchmark.

MCP serverMCP · TypeScript · LanceDB

przm Voice

Live

A personality layer that keeps an AI assistant's voice coherent and evolving across sessions instead of resetting every chat.

MCP serverMCP · TypeScript

Cortex

Beta

Permission-aware company knowledge for AI agents. ACLs enforced at the database layer, deployable on-prem when data cannot leave.

MCP + On-premMCP · PostgreSQL RLS

SnapTab

Live

Receipt scanning with vision models in a shipped consumer app: photo in, itemized split out, on both app stores.

iOS + AndroidGemini Vision · React Native

Dungeon Diary

Live

AI world building, NPC generation, and portrait generation for tabletop campaigns, in production for real Dungeon Masters.

WebClaude · Next.js · PostgreSQL

AI agents built the page you are reading

This site was assembled by a squad of AI coding agents working from written briefs, reviewed and merged by the studio. This is the actual run log.

run r-20260711-081118 · missions/studio-site

The same discipline applies to client work: agents do the labor, written briefs and test gates keep them honest, and a person who knows the codebase reviews every line before it ships.

Nexus runs the studio

Nexus is the agent platform the studio built to run its own operations. It is where everything on this page stops being theory.

Agents on Nexus handle real work: qualifying leads, drafting outreach, watching client health, standing up infrastructure for new projects. They do it inside a structure that assumes AI will sometimes be wrong.

That structure is the product. Agents do not improvise; they emit declared events, ask a policy engine before acting, and queue anything sensitive for a person. The studio trusts its own business to this, which is a higher bar than trusting a demo.

14 packages241 passing testsAgent-built under review gates

The same platform thinking, sized to fit, is what client AI work gets.

L01

Event spine

Every action an agent takes is a typed event with a schema and a trace. Nothing happens off the record, and every business object can be replayed from its history.

EventsSchema registryTracing

L02

Policy and autonomy

Every side effect clears a policy check first. Agents earn autonomy one level at a time, anything a client would see waits for human sign-off until the agent has earned better, and kill switches stay hot.

Policy engineAutonomy ladderKill switches

L03

Agent mesh

Specialized agents for lead handling, client retention, and business insight cooperate over declared contracts. An agent that emits something it did not declare gets flagged, not forgiven.

Lead-genRetentionInsight

L04

Genesis factory

Agents that provision real infrastructure for new projects: repository, deployment, environments, end to end, with release control on top.

RepoDeployEnvironments

L05

Attention queue

Decisions that need human judgment land in one queue with context, options, and a recommendation. The whole platform is observable on a live dashboard.

ApprovalsDashboardLive data

From hype to habit

Start small, measure honestly, and scale only what works.

01

Audit

A working session over your actual workflows and data. Sort the AI ideas into three piles: pays for itself, worth a pilot, and hype. You keep the writeup either way.

02

Pilot

One workflow, working software in weeks. A pilot is code in your stack with your data, not a slide deck about possibilities.

03

Harden

The pilot grows evals, guardrails, budgets, and approval gates before it touches customers or production data at scale.

04

Enable

Your team learns to run and extend it: AI tooling, agent workflows, review discipline. The goal is not needing the studio for every change.

Not sure which step you are on? A short conversation usually makes it clear.

Talk about AI