AI product studio
NCM Labs helps teams turn LLMs and agents into products that actually work: prototypes in days, then hardened with evaluation gates, supervision and full traceability.
Live demo · Agent console
Pick a job, inject a problem, press run. Every step shows the tool it called, how sure it was and which rule fired. When the stakes rise, the agents stop and ask you.
Request
“Close September for Example Studio SRL: collect every receipt and invoice, match the payments and hand it all to our accountant.”
Scripted replay with synthetic data. Same method we use on real systems.
Run trace
0 stepsHuman decision required
Live artifact
Handoff pack| Document | Amount | Status |
|---|---|---|
| INV-0912 · Hosting | 49.00 EUR | |
| RCPT-0915 · Supplies | 212.40 RON | |
| INV-0918 · Contractor | 3,200.00 RON | |
| RCPT-0921 · Fuel | 310.75 RON | |
| INV-0926 · Software | 19.99 EUR | |
| Bank · 24 Sep transfer | 1,450.00 RON |
bracket.step · exported
Run complete
Services · What we build with you
LLMs and agents change how fast software gets built and what it can do. We bring both, plus the engineering that keeps them honest.
From idea to a product people use: LLM features and agent-powered apps, designed around the moments where the model can be wrong.
A working product, not a demo.
Repetitive, document-heavy or operational work handed to supervised agents with budgets, confidence thresholds and a human sign-off where it matters.
Hours back every week, with an audit trail.
Choosing, adapting and serving the right model for your task, then proving it with held-out evaluations against explicit baselines.
Decisions backed by measurements.
Connecting assistants to CAD, fabrication and embedded systems through inspectable tools and validation between intent and geometry.
Conversation becomes checked, physical output.
Process · Idea to production
Agents compress the build. Gates, evaluations and human checkpoints are what make the result safe to rely on.
We map the workflow, where a model adds value, and where it must never act alone. You get a scoped plan and success metrics.
Agents and LLMs let us put a working slice in your hands fast, on your real inputs, so decisions come from use rather than slides.
Evaluation sets, confidence thresholds, budgets and escalation paths turn the prototype into something you can trust under load.
Deployed, traced and measured. Every decision, cost and fallback leaves evidence your team can inspect and improve.
Work · Built in-house and in the open
A programming language for LLM systems, bridges from agents into CAD and CNC, and products we run ourselves. Open-source where we can, measured everywhere.
A programming language for oracle-augmented computation. LLM calls are oracle queries with their own type: uncertainty is tracked at compile time, deterministic code is structurally separated from stochastic code, and failure handling is declared, not bolted on.
Uncertain results become explicit, handled values.
task classify_intent
needs message: Text
gives Intent
do
result = classify message into ["buy", "support", "cancel", "other"]
when result.sure(above: 0.85) -> give result when result.sure -> give result with flag("low-confidence") when result.unsure -> give ask_for_clarification(message) else -> escalate to humangive
“support” returned to the caller.
A portal that collects monthly accounting evidence, resolves missing documents and payment exceptions, and prepares a traceable handoff for the accountant. Extraction is measured against an evaluation corpus before it is trusted.
Month-end becomes a checklist, not a chase.
One structured command surface an agent can drive to manage compute, storage, networking and GPUs, with local capacity first and cloud overflow when needed. Every command returns status, data and explicit next steps.
Infrastructure an agent can operate, inside guard rails.
An MCP server that lets an AI assistant inspect and manipulate a live Autodesk Fusion 360 session: sketches, features, measurements, interference checks, assemblies and joints.
Conversation becomes inspectable geometry.
you › L-bracket, 40 × 40 mm, 3 mm, two M5 holes
FlatCAM post-processors that emit Smoothieware G-code for the Makera Carvera Air: PCB isolation milling, drilling and laser, with a two-pass auto-probe toolchange for multi-tool jobs.
Board design to machined PCB without hand-editing G-code.
Also in the lab · Private R&D
Reproducible loops for adapting models to specific tasks: data curation, training, serving and held-out evaluation treated as one system.
Sensor-rich devices and distributed control, with firmware, feedback and safety designed together from the start.
Approach · Four invariants
Anyone can wire an LLM to an API. A system that behaves honestly under uncertainty, failure and real consequences takes a different set of decisions. These are ours.
Stochastic work should be visible in the architecture, not disguised as a normal function call.
Reasoning can explore. A typed or physical action must pass an explicit boundary first.
When confidence or consequence demands it, the system escalates instead of improvising.
Decisions, costs, fallbacks and execution paths leave evidence an operator can inspect.
Contact · Start a conversation
Tell us what you want to automate or build. We reply with honest next steps, even if that means "you don't need an LLM for this".
We can't find the internet
Attempting to reconnect
Something went wrong!
Attempting to reconnect