Consulting · Cryptuon
Agents with domain knowledge.
Built like production software.
Most "AI agent" projects are LLM prompts over an API. We build the other kind — typed domain models, deterministic control loops, risk invariants, MCP tool surfaces, and the evaluation harness that proves they work. polybot is the public, MIT-licensed showcase; client work goes deeper.
What we build
Domain agents, not prompts
Custom agent cores with typed tools, memory, planning, and evaluation harnesses. Built on Claude 4.x or your model of choice.
Expose your platform to agents
Design and ship Model Context Protocol servers for your product — turning any SaaS or internal platform into an agent target.
Prediction, prop, on-chain
Strategy research, execution plumbing, venue integrations, risk infrastructure. polybot-grade hardening, on your markets.
Proving the agent works
Eval harnesses, regression suites, shadow-deploy infrastructure. No more "it looked good in the demo".
Caching, routing, budgets
Prompt caching, model routing, per-tenant budgets, observability. We take your token bill seriously.
Ship the notebook
Take a research prototype, harden it into a production service with monitoring, tests, and deployment.
Case study: polybot
polybot is the kind of system we build for clients — open-sourced so you can read the code instead of taking our word for it.
- Typed domain model. Pydantic everywhere. Signals, orders, positions, markets — every handoff is type-checked.
- MCP server as a first-class interface. Not bolted on; designed alongside the core so tool semantics match the domain.
- Async service mesh. Scanner, executor, analytics communicate via NNG pub/sub — decoupled for independent scaling and failure isolation.
- Plugin systems. AI models and venues are behind interfaces. Swapping Claude for GPT, or adding Drift, is a contained change.
- Risk as infrastructure. Caps, limits, shadow mode, approval gates — enforced in the platform, not each strategy.
The repository is github.com/cryptuon/polybot. Read the code, read the docs, and if that's the bar you want for your agent system, book a call.
How we work
- Week 0 — discovery. A paid one-week deep dive. You get a written architecture memo and a shippable PR even if we don't continue.
- Two-week sprints. Weekly demos, PR-based deliverables, your repo or ours (your choice).
- Handover built-in. Documentation, runbooks, and pairing with your team from day one. No "you need us forever" moat.
FAQ
What does Cryptuon build?
Production AI agent systems with domain knowledge — trading, on-chain execution, market microstructure, MCP servers, LLM orchestration. We ship code you own, not SaaS lock-in.
Do you take equity or only cash?
Both. For early-stage teams with a real wedge, we do a mix. For funded teams and enterprises, cash engagements on weekly or monthly retainers.
How fast can you start?
Typical kickoff is 1–2 weeks from intro call to first PR. We front-load a discovery week, then ship in 2-week sprints.
Is polybot your only work?
polybot is our open-source showcase. Client work is under NDA, but spans prediction markets, crypto derivatives market-making, and agent infra for fintechs.
Let's build the agent.
30 minutes. Bring your hardest question. Leave with an architecture sketch.