$ humano-lab --status
AI is powerful.
Production is harder.
An AI engineering lab. Cost reduction, evals, and agent and workflow architecture — the systems work most teams can't get out of prototype. We ship in weeks, not quarters.
// for engineering leaders and founders whose company…
- has token bills growing faster than revenue
- has an AI prototype that works in demo but breaks in production
- has shipped 3 PoCs and 0 production AI features
- wants to add AI features or capabilities to an existing product
- wants to automate operations with AI agents, end-to-end
- can't decide between agent and workflow architecture
What we work on
Three AI engagements, each scoped tight. Pick the one that matches the pain you actually have.
-
01 ai-cost-reducer booking Your LLM bills are growing faster than your revenue. We audit your inference path end-to-end, find the waste, and bring costs back to earth.
delivers/
- Find the 40-70% you're overpaying
- Switch only what changes the bill, nothing else
- Guardrails so it doesn't blow up again next quarter
based on: production AI runtimes we have shipped with real inference budgets
-
02 evals-and-reliability booking Your AI feature works in the demo and breaks in production. We instrument it, design the evals, and ship the changes that turn 'works most of the time' into a metric you can defend.
delivers/
- Catch quality regressions before your users do
- Know within minutes when a model change breaks something
- Stop manually re-running prompts to check if it still works
based on: reliability work for AI products and country-scale software (millions of users)
-
03 agents-and-workflows booking Should it be an agent or a workflow? We design and ship the answer. Production-grade architecture for AI features that do real work — tool use, multi-step reasoning, integration with the rest of your stack.
delivers/
- We decide for you: agent or workflow, with reasons
- Real integration with your CRM, ERP, and internal APIs
- Humans intervene at the right step, not random ones
based on: the Rust-based local-first agent runtime we shipped (MCP-style, pre-MCP)
How we work
Engagements built around outcomes, not hours.
We understand the problem, the constraints, and what "done" actually looks like. We say no early if we are not the right fit.
A written proposal: deliverables, timeline, price, and the assumptions behind them. No surprises.
We build it, ship it, and stay involved long enough to measure that it actually moved the metric you cared about.
Our work
AI systems, shipped in production.
AI agent runtimes, enterprise agent platforms, and MCP-era tooling — designed and shipped in production.
-
01 local-first-agent-runtime in production We designed and built a privacy-preserving, local-first AI agent runtime in Rust, shipped as a cross-platform desktop app. The runtime, the tool/protocol layer, and OAuth-based capabilities — an MCP-like architecture, shipped before MCP became a public standard.
notable: shipped MCP-style architecture before MCP was a public standard
-
02 enterprise-agent-platform early product We bootstrapped an enterprise AI-agent platform from zero-state into an early product. It certifies, integrates, and measures the ROI of AI agents across business operations — ERP, CRM, messaging, and more.
notable: certified, integrated, and ROI-measured agents across business operations
-
03 agent-server-manager open-source We designed and shipped an installer and manager for AI agent / MCP servers, packaged as a cross-platform desktop application, a CLI, and a library. Used by AI developers across the ecosystem.
notable: used by AI developers across the ecosystem
-
04 agentic-city-sim shipped We built the AI-agent layer for a city simulation. Agentic NPCs reason over on-chain state, integrated end-to-end with the underlying Layer-2 network and AI services.
notable: agentic NPCs reasoning over on-chain state, end-to-end integration
positioning
Most AI features are a prompt and a prayer. We are the lab you hire instead of praying.
What we believe
Opinions we are willing to defend in production.
Most "AI features" today are a prompt and a prayer. We treat AI like the systems engineering problem it actually is: evals, instrumentation, and architecture you can defend.
Slow software costs money, attention, and user trust. When the inner loop matters, we reach for Rust and design for the budget the system actually has.
Some workloads belong on your servers, some on the user's machine, some neither. We pick where compute lives based on privacy, latency, and cost. Not on what is fashionable.
We build systems that make humans faster and sharper. Not systems that pretend to replace the judgment a human still has to take responsibility for.
FAQ
Things people usually want to know up front.
01 How do you engage: fixed price, retainer, or hourly?
Fixed-price for scoped work whenever possible: both sides know what we are signing up for. Retainer for ongoing relationships. We avoid hourly because it punishes the engineer for being fast.
02 What time zones do you work in?
We are based in Chile (UTC−3 / UTC−4) and work comfortably with teams in the Americas, Europe, and most of EMEA. We are async-first and good at it.
03 How fast can you start?
Usually within two to three weeks of agreeing on scope. We take a small number of projects at a time, so the slot you book is the slot you get.
04 What if my AI problem doesn't fit one of your three services?
Talk to us anyway. The three services are the work we go deepest on, but most AI engagements have edges that touch all three. If we are not the right fit, we will tell you on the first call.
05 Do you only work remotely?
Yes. We are remote-first and have shipped that way for years. We can travel for kickoffs or critical reviews when it genuinely matters.
06 How big are your typical projects?
From two-week focused audits (think LLM cost reductions, eval design) to multi-month builds (production agent platforms, end-to-end AI features). We are happy to scope something smaller if it is the right starting point.
07 Do you write Rust? Do non-AI engineering work?
Yes to Rust — we use it when latency, safety, or memory pressure matter (the local-first AI agent runtime in our past work is in Rust). No to non-AI as a standalone offering: we focus on AI engineering, and bring Rust and infrastructure depth to those engagements where it serves the AI work.
08 Who owns the IP and code we ship together?
You do. Anything we ship under the engagement — code, models, prompts, evals, documentation — is yours. We retain ownership only of our generic methodology and reusable internal tooling. Clean separation, no surprises.
09 Will you sign an NDA?
Yes. We sign a mutual NDA before scoping or before seeing your code. We do not need one to listen on the first discovery call, but we will always sign before anything specific is exchanged.
10 What if we're not happy with the work?
Fixed-bid means we cover overruns. If we agreed the engagement would move metric X to Y and it doesn't, we keep working until it does, without a second invoice. We scope carefully so this rarely happens, but the commitment is the commitment.
contact
Tell us what's stuck
in your stack.
We take a small number of projects per quarter. The slot you book is the slot you get.