Broentech

Broentech AI-native software, built end-to-end by one engineer. 18 years across IoT, energy, utilities, fintech, maritime, sustainability.

An avatar's face isn't animation. It's a running state machine.Most chatbot faces are choreography. A library of canned ...
19/08/2026

An avatar's face isn't animation. It's a running state machine.

Most chatbot faces are choreography. A library of canned expressions, triggered by keywords, played back like a GIF. Pleasant. Decorative. Disconnected from what the model actually understands.

ASMI is built the other way around. The name is the architecture: Avatar State Machine Interface. Sentiment and intent from the conversation drive expression transitions in real time, so the face is an output of the reasoning, not a costume on top of it.

We call this layer the Expression State Machine. It sits between the model and the face, and its state can be serialized, inspected, replayed, and resumed across sessions. You can debug a feeling the same way you debug a request 💡

Decorative faces age badly. Once the novelty wears off, users notice the seams. A state machine doesn't have seams. It has transitions.

The industry settled on a text box and called it a conversation.We built the alternative, and found out what was actuall...
17/08/2026

The industry settled on a text box and called it a conversation.

We built the alternative, and found out what was actually missing.

Most teams assume a well-written chatbot reply is enough to create a human connection. The interface gets treated as a cosmetic wrapper. But the interface IS the interaction. When you funnel a sophisticated model down to a flat text bubble, you restrict how visitors process and trust what they're reading.

So when we built ASMI, we decoupled the interface from the underlying LLM. Instead of decorative animation or pre-rendered choreography, we designed an expression state machine that reads model output and classifies sentiment and intent in real time. The face shifts as the response is generated, reactive to the actual conversation logic underneath 💡

We build this way in Horten because sixteen years of shipping production AI has taught us one thing: trust is established in the subtle, structural layers.

CX teams pouring budget into model quality while leaving the interface untouched are bottlenecking the very intelligence they built.

A smart model behind a still face is a much narrower channel than it has to be.Text-only chat throws away most of what a...
14/08/2026

A smart model behind a still face is a much narrower channel than it has to be.

Text-only chat throws away most of what a conversation carries. Tone, hesitation, warmth, surprise. The model understood all of it. The interface just refused to show it.

ASMI closes that gap with an Expression State Machine. It sits next to your LLM, reads the output as it streams, and drives the avatar's face in real time based on what the model just understood about the visitor.

Not animation. Not choreography. State.

When the model softens, the face softens. When it clarifies, the brow does the work a punctuation mark can't. Rip the Expression State Machine out and the product collapses into another chatbot. It's structural 💡

Two things worth saying out loud: ASMI is model-agnostic, so bring your own LLM. And the runtime is MIT-licensed, so the avatar keeps running on your stack even if our backend is unreachable.

Sixteen years of shipping production AI from Horten taught us that the interface is where most AI products quietly fail. The face is where we decided to stop failing.

Most avatars break in production because nobody tested them as a system. They tested the script.We built ASMI around a t...
12/08/2026

Most avatars break in production because nobody tested them as a system. They tested the script.

We built ASMI around a three-stage standard. Design. Test. Deploy. In that order, every time 💡

Design happens in the visual editor. Lock baseline identity, generate expression frames, configure wardrobe presets, decide which looks fire under which conditions. All before a line of code.

Test happens in the Simulator. Run the avatar against the live state machine, watch every sentiment classification and state transition in real time, and record the canonical happy-path as your ground truth.

Deploy happens through the MCP server. Your coding agent installs the open runtime, writes the glue, and verifies against the recorded happy-path before anything reaches a visitor.

The point is sequence. Identity locked before behavior is tested. Behavior verified before code is written. Code checked against a known-good recording before it ships.

We call it the DTD Loop. Which stage does your current avatar workflow skip?

Most teams come to us having already accepted a bad trade: rent a vendor SDK with per-seat fees and runtime lock-in, or ...
10/08/2026

Most teams come to us having already accepted a bad trade: rent a vendor SDK with per-seat fees and runtime lock-in, or hand-roll the face, state machine, and expression layer from scratch. Both look expensive. So the interface gets pushed to next quarter.

That's the wrong diagnosis. The design problem and the implementation problem are not the same problem.

In ASMI, you design the avatar's identity, expressions, knowledge, and behavior in a visual editor. What you publish is a blueprint, not a service.

The runtime lives in your codebase. MIT-licensed packages, your LLM keys, your hosting. Bring your own model, swap it when a better one ships. Your avatar keeps answering even if broen.tech goes offline tomorrow 💡

We've spent 16+ years shipping production AI from Horten. A blueprint is not a service. Your stack stays yours.

Most avatar vendors want to own your model. We don't 💡ASMI sits next to your LLM. It doesn't replace it, doesn't wrap it...
07/08/2026

Most avatar vendors want to own your model. We don't 💡

ASMI sits next to your LLM. It doesn't replace it, doesn't wrap it, doesn't route around it.

Your model generates the response. ASMI reads that output, classifies sentiment and intent, and drives the avatar's expression from there. Your LLM provider, your hosting, your data, your prompts. Untouched.

That's what "Bring Your Own Model" actually means at the architectural level. The interface layer is decoupled from the intelligence layer. The face on your site reacts to your model's output, not ours.

Avatar products that own the LLM also own your roadmap. Switch providers, renegotiate pricing, change prompts, and the avatar breaks. Vendor lock-in dressed up as a feature.

We built ASMI the other way because that's the only way it survives production. Swap GPT for Claude for a fine-tuned local model on a Tuesday afternoon, and the avatar keeps working. Same identity, same expressions, new brain.

If your current avatar vendor can't tell you exactly where their system ends and your LLM begins, that's the answer.

A smart model behind a dead text box is a much narrower channel than it has to be.We've watched teams misdiagnose this a...
05/08/2026

A smart model behind a dead text box is a much narrower channel than it has to be.

We've watched teams misdiagnose this all year. The assumption goes: smarter model means better experience. So the budget pours into prompt engineering, fine-tuning, model selection. The interface stays a blinking cursor in a white box.

Then they ship, and the feedback comes back cold.

The bottleneck almost never lives in the model. It lives in the channel the model has to speak through. A reply with no face, no expression shift, no visible acknowledgment of tone reads as mechanical no matter how good the reasoning underneath is. Swap GPT for Claude for your own fine-tuned model. The visitor still feels like they're filling out a form.

This is what we built ASMI to fix. Not a smarter brain. A wider channel. The face reacts to sentiment and intent in real time, expression state runs as system logic rather than animation, and the model your team already picked finally starts feeling responsive instead of robotic.

If your conversational AI plan for next quarter is entirely about the model, you're optimizing the half that's already working. Which half of your stack actually needs the next investment?

The LLM gave a perfect answer. The visitor still bounced.We see this almost every week in client work. A SaaS team upgra...
03/08/2026

The LLM gave a perfect answer. The visitor still bounced.

We see this almost every week in client work. A SaaS team upgrades to a better frontier model, expects engagement to climb, then watches the bounce rate sit exactly where it was. The model wasn't the bottleneck. The interface was.

Text-in, text-out chat is a thin surface for something that's actually reading intent, hesitation, frustration, curiosity. The intelligence is there. The visitor just can't feel it.

That's what ASMI does. It reads the LLM's output, classifies sentiment and intent in real time, and drives a reactive face that shifts expression as the conversation moves. Confusion softens. A direct answer lands with directness. Reassurance looks like reassurance 💡

Expression State is system logic, not animation. It's bound to what the model is actually saying, not scripted choreography on a loop. And because it's model-agnostic, the LLM stays yours. Swap GPT for Claude for an in-house model. The face keeps reading sentiment the same way.

After 16 years of shipping production AI from Horten, this is the gap we see most often: smart systems that feel cold because nothing translates the intelligence into something a human reads in 200 milliseconds.

Upgrade the model if you want. But if visitors are bouncing on correct answers, the model isn't the conversation you need to have.

When an avatar gives a wrong answer, the first instinct is almost always the same: add more knowledge items. We've watch...
31/07/2026

When an avatar gives a wrong answer, the first instinct is almost always the same: add more knowledge items. We've watched this play out across every industry running ASMI in production. It almost never fixes the problem. Usually it makes it worse.

Knowledge volume isn't the accuracy problem. Retrieval precision is.

Bulk-import 800 pages of overlapping FAQs, product specs, and legacy help docs, and you're not closing gaps. You're handing the retrieval pass three slightly different answers to the same question. The model picks one. Sometimes the wrong one. Sometimes it blends two and contradicts itself in the same sentence.

This is why ASMI's knowledge layer is built around an item lifecycle, not a dump 💡 Every item has a state: draft, published, archived. Retrieval only sees published. Page-aware retrieval narrows things further, so a product page question doesn't pull in a support article that shares five keywords. And the crawl workflow accepts items one by one on purpose. Slower than a bulk import. That's the feature.

After 16+ years of shipping production AI from Horten, the pattern in CX teams is consistent: they treat the knowledge base like a warehouse and the retrieval layer like a search bar. Both assumptions are wrong. Curation beats volume, every time.

If your avatar is giving contradictory answers, don't add. Archive.

A client shipped their ASMI avatar in January. Six months later: same outfit, same look, every single day.Not because th...
29/07/2026

A client shipped their ASMI avatar in January. Six months later: same outfit, same look, every single day.

Not because they forgot. They assumed changing it meant rebuilding from scratch. We see this pattern constantly. Builders fuse identity and appearance at launch, and from that moment the avatar is frozen.

That assumption is the bug 💡

In ASMI, identity is locked once. Appearance is a scheduled layer on top. Wardrobe presets decouple styling from the underlying avatar. Same face, same trained knowledge, same expression state logic, all untouched. What rotates is the look: a daily default, a seasonal variant, an event outfit for a launch or a holiday week.

Three activation modes do the work. Default for the everyday look. Dynamic when context shifts. Forced when you want a specific outfit live for a specific window, scheduled on the calendar. No rebuild. No retraining. No touching the expression system that drives sentiment and intent in real time.

Same face on Monday. Different jacket on Friday. Holiday look in December, queued in October.

If your avatar has been wearing the same thing since launch, that's not a design choice. That's an architecture you can change this week.

Adresse

Damgata 53
Åsgårdstrand
3179

Nettsted

Varslinger

Vær den første som vet og la oss sende deg en e-post når Broentech legger inn nyheter og kampanjer. Din e-postadresse vil ikke bli brukt til noe annet formål, og du kan når som helst melde deg av.

Kontakt Bedriften

Send en melding til Broentech:

Snarveier

Del

Type