01  ·  AI & MACHINE LEARNING

AI agent development that reaches production

Most AI work stalls between a convincing demo and a system anyone can depend on. The gap is rarely the model. It is evaluation, latency, error handling and the unglamorous work of making a probabilistic component behave inside a deterministic product. Four of the systems below are live right now and you can open every one of them.

Book a scoping callFree, and it ends with a range or a straight no.

What we build

Five shapes of work, all of them shipped rather than theorised.

LLM and agent systems

Multi-step agents with tool calling, orchestration and a defined failure mode for every step. Built so a wrong answer is caught rather than served.

RAG systems that cite their sources

Retrieval that keeps the link between a claim and the document it came from, so a conclusion can be defended rather than trusted. Axiom runs to roughly 50 sources per report on this pattern.

Real-time voice agents

Streaming speech-to-text, turn detection and synthesis that starts speaking before generation finishes. The assistant on this site is one, and you can interrupt it mid-sentence.

Document intelligence

Extraction and classification pipelines over messy real-world documents, with a confidence threshold that routes the uncertain cases to a person instead of guessing.

Predictive modelling and computer vision

Forecasting, classification and vision models, with the evaluation harness built at the same time as the model rather than after it.

Why most AI agents never reach production

Four failure modes account for nearly all of it, and none of them is about picking the right model.

Evaluation is the part most vendors skip

Without a regression suite over prompts, a change that fixes one case silently breaks four others. We build the harness first, so quality is measured rather than asserted.

Model deprecation will break your application

Providers retire model ids on a rolling schedule and do not keep them alive. This site runs a health check against its own model id for exactly that reason, and it is an early warning rather than an outage.

Latency is a product decision, not a tuning pass

A voice agent that takes three seconds to answer is a different product from one that takes eight hundred milliseconds. That has to be designed in at the architecture stage.

Nobody owns it after handover

Every build leaves with documentation, CI, observability and a team that can run it without us. That is the definition of done.

How an engagement runs

Paid discovery sprint first, then a fixed-scope build. An MVP is typically four to eight weeks; larger platforms run in six-week phases with a working demo at the end of each. The success metric is agreed before the first sprint and reported against in every review.

4 of these are live right now

Not a portfolio of screenshots — every one of these opens.

  • Maiku AI

    Real-time interview copilot

    A desktop overlay that transcribes the interview live with Groq Whisper and drafts structured answers with LLaMA while the call runs on Zoom, Meet or Teams. The window is registered with the Windows Display Affinity API, so it is physically absent from captured frames rather than hidden behind another window. Core is open source and runs on the user's own Groq key.

    maiku.app
  • ArchitectXpert

    AI floor plan generator

    A generator that takes plot size, floors, bedroom and bathroom counts, city and house style, then produces a dimensioned architectural plan with room labels, door placements and a total area breakdown. Output downloads as PNG or DXF for AutoCAD, so an architect or contractor can pick the drawing up and keep working in it.

    architectxpert.tech
  • Axiom

    Private research instrument

    A research workspace that turns a question into a defensible trail: it reads sources, separates claims from reasoning, and keeps the link between them. Reports stay in a private library with a source collection per project, and the workspace shows how much research context a report consumed.

    axiom-agent-three.vercel.app
  • Cartesia Assistant

    Low-latency voice assistant

    A call-shaped assistant built on Cartesia speech synthesis with streaming turn-taking, so a reply starts speaking while the rest is still being generated. A live waveform shows who holds the turn, and users choose from a roster of voices before the call starts.

    cartesia-assistant.vercel.app

What we build on

  • PyTorch
  • LangGraph
  • Vertex AI

Halyx is not religious about the stack — it picks what your team can maintain — but these are the defaults it is fastest and safest in.

Questions we get asked first

How much does it cost to build an AI agent?
It depends on how much of the work is retrieval, how much is integration, and whether a person stays in the loop — but that is not a useful answer on its own. The first conversation is a free scoping call that ends with either a range and a proposed first step, or a straight answer that Halyx is not the right fit.
How long before we see something working?
A discovery sprint produces a decision, not a slide deck. An MVP typically lands in four to eight weeks. Larger platforms run in six-week phases, each ending in a demo you can use.
Who owns the code?
You do. Every build ships with documentation, CI and observability, and the explicit goal is that your team can run it without us.
Can you work alongside our existing engineering team?
Yes, and it is often the better shape. We have handed systems over to client engineers who owned them from that point on. Where in-house is genuinely the right call, we will say so.
What uptime or SLA do you guarantee?
Halyx publishes no standard uptime figure, because a number promised on a website before anyone has seen the system is worth nothing. Availability targets and support hours are scoped per engagement and written into the contract.

The other four practices

Talk to the people who would build it

The first conversation is a scoping call, not a pitch: what you are trying to move, what already exists, what the constraint is. It ends with either a range and a proposed first step, or a straight answer that Halyx is not the right fit.