Broentech
ON THE HORIZONWhat's coming to ASMI

The next few releases.

A short, honest roadmap. Some of this is in the codebase already. Some of it is still in feasibility studies. Listed here so you can see where the product is heading — and so you can push back on priorities before the commits land.

In development

Actively being built. Expect this in a near-term release.

In research

Feasibility studies and prototypes. Might ship, might reshape.

Planned

On the list and shaped up enough to describe. Timing depends on what the first two columns teach me.

In developmentHosted intelligence layer

Secure, compliant endpoints for agentic avatars.

Today ASMI is bring-your-own-LLM: you point the runtime at a provider and the avatar talks. That works, but it pushes the compliance burden onto every host site. The next layer is a Broentech-operated endpoint — authenticated, rate-limited, audit-logged — that lets an avatar call third-party systems on a visitor's behalf without you having to wire each integration yourself. Booking, support ticket creation, transactional queries. You get the agentic behaviour; we carry the security posture.

In researchOn-device intelligence

A model small enough to live in the browser.

The most private avatar is the one that never talks to a server at all. Small open-weight models are improving fast and WebGPU is landing on every modern device. The bet is that within a couple of years a model small enough to ship with the page will be capable enough to drive the conversation a mid-site concierge needs to have — and a visitor's words never leaving their own machine is a strong value proposition. I'm testing and evaluating now so we're ready when the capacity catches up to the ambition.

In researchOn-page action

An avatar that can actually do things on the page.

Once the avatar runs locally, the next question is what it's allowed to touch — filling a form, navigating a flow, completing a booking without a human click. That capability is useful. It's also a cybersecurity category all its own. The research right now is less about "can we" and more about "what's safe enough to ship, to whom, and under what constraints." Guard-rails before capabilities.

PlannedSmooth animation

Vectors instead of keyframes.

The face today is a set of expression stills — crisp, but with no in-between. The plan is to convert avatars to vector-based representations that can be animated smoothly through expression changes and state transitions. A smile doesn't cut; it spreads. More life at a lower per-expression cost than rendering every micro-state separately.

PlannedTTS integration

Let the avatar speak.

Text is enough for most things, but not for everything — accessibility, ambient use, reading-while-driving, a kid who hasn't learned to read yet. TTS integration is on the roadmap, and pairs naturally with vectorized avatars whose mouths can move in sync with the audio.

In researchStyle range

Beyond the illustrated look.

Every avatar today ships in one illustrated style. It's on-brand for a lot of sites, but not for everyone. Anime, semi-realistic, photorealistic, line-art, watercolour — each opens a different set of hosts and use cases. I'm evaluating generation models and consistency clauses that would let a new style drop into the editor as a first-class choice, alongside the current one.

Timing is deliberately absent from this page. I'd rather ship each item when it's ready than ship a date that slips. If a particular item matters to you, say so — prioritisation responds to what people actually want.

NEXT STEP

Build the front door your product deserves.

Start with an avatar, test it in the simulator, add the knowledge it needs, and share the result when it is ready. Examples in the gallery, full technical reference on the developer page.

A product by Broentech Sentinel.

Broentech Sentinel