Welcome to the ellix.ai Blog

We're building AI tools that actually work for businesses — not demos, not toys, but production-ready products that integrate into your workflow and stay reliable under real load.

This is the first post on the ellix.ai blog. We're going to use this space to document what we're building, the technical decisions behind it, and the lessons we're collecting as we ship. No thought leadership. No vague predictions about "the future of AI." Just concrete, honest writing from a team that builds things.

Why We Started ellix.ai

The AI tooling market is dominated by two extremes: toy integrations that impress in demos but fail in production, and enterprise platforms that require a six-month procurement cycle and a dedicated ML engineer to operate. The businesses caught between those extremes — the 10- to 200-person SaaS companies, the regional e-commerce brands, the agencies — are being systematically underserved.

These businesses have real problems. They're drowning in repetitive support tickets. Their sales teams can't answer product questions fast enough to capture buying intent. Their onboarding flows leak users because questions go unanswered at 11pm on a Tuesday.

AI can solve all of these problems. The technology is ready. What's missing is tooling that's actually designed for this tier of company — fast to set up, affordable at their scale, and reliable enough that they don't need to babysit it.

That's the gap we're filling.

Our mission: ship AI products that work without an ML team, work on day one, and keep working as your business grows.

What aiassist.chat Does

aiassist.chat is an embeddable AI chat widget that answers questions about your product using your own content as the knowledge source. It's not a general-purpose chatbot trained on the internet. It's a grounded assistant that knows exactly what you've told it, and nothing more.

The setup takes about ten minutes:

Add one <script> tag to your site
Upload your documentation — paste a URL, upload a PDF, or let our crawler index your public pages
Your AI assistant goes live

When a visitor asks a question, the system retrieves the most relevant chunks from your knowledge base and passes them to the language model as context. The answer is grounded in your actual content, with source citations the visitor can verify. If the knowledge base doesn't contain a relevant answer, the AI says so and routes to human support.

What Makes Our Approach Different

Most embeddable chat widgets are thin wrappers over a general-purpose LLM with a custom system prompt. They work until someone asks a product-specific question that the LLM's training data doesn't cover — then they hallucinate.

We built differently. The core of aiassist.chat is a RAG (Retrieval Augmented Generation) pipeline:

Every document you upload gets chunked and embedded using text-embedding-3-small (1536 dimensions)
Vectors are stored in a per-tenant Qdrant collection — complete data isolation, no cross-tenant leakage
At query time, we embed the user message, run nearest-neighbor search, retrieve the top-K chunks, and send them as context to Claude or GPT-4o

This architecture means the AI's answers are only as wrong as your documentation. Stale doc? Update it. New feature? Add it. No retraining, no model deployment, no ML expertise required.

The Technical Stack

We made opinionated choices on the stack and we'll defend all of them:

Rust + Axum for the backend. Two services — website-core handles billing, auth, and tenant management; ai-engine handles chat, embeddings, RAG, and streaming. The Rust compiler prevents entire categories of bugs that you'd only discover at 2am in a Node.js or Python service. The memory footprint is tiny — website-core runs under 30MB at idle, which matters when you're running many tenants on a small fleet.

Per-tenant Qdrant collections. One collection per tenant, named tenant_{uuid}. Isolation at the storage layer means one tenant's heavy indexing workload cannot degrade search latency for others.

Redis for rate limiting. Sliding window counters using INCR + EXPIRE — atomic, fast, and correct under concurrent load in a way that in-process counters are not.

React 19 + TanStack Start for the dashboard and this website. Type-safe routing, colocated loaders, SSR. We'll write more about the TanStack Start experience in a later post — the short version is that the type-safety story is worth the ecosystem immaturity.

Claude and OpenAI for models. We support both. Tenants can configure which model they want for their assistant. Claude 3 Haiku handles the majority of queries cheaply; Claude 3.5 Sonnet is available for teams that need higher quality on complex questions.

Embed aiassist.chat Right Now

If you want to try it today, the setup is genuinely one script tag:

<script
  src="https://cdn.aiassist.chat/widget.js"
  data-site-id="YOUR_SITE_ID"
  data-position="bottom-right"
  defer
></script>

Get your site ID from the aiassist.chat dashboard, upload a few docs, and your AI assistant is live. The free tier covers 100 conversations per month — enough to evaluate whether it's solving real problems for your visitors before you commit to a paid plan.

What's Coming Next

The immediate roadmap:

Analytics dashboard — per-question performance, confidence scores, unanswered question tracking, and knowledge base gap reports
Proactive triggers — show the widget based on time-on-page, scroll depth, or exit intent rather than only waiting for visitors to initiate
Human handoff improvements — richer conversation context transfer when escalating to your support team, with CRM integration
Multi-site management — a single aiassist.chat account managing assistants across multiple domains, with shared or isolated knowledge bases per site

We're shipping fast. If you have a use case you want to see supported, the team reads every email at hello@ellix.ai.

Following Along

This blog will cover the technical decisions, the product experiments, and the lessons we collect as we build. If you're building AI-powered products or just want to understand what good AI tooling actually looks like under the hood, subscribe to the RSS feed or follow us on X.

Building in public is both nerve-wracking and clarifying. We'd rather be honest about our tradeoffs than pretend everything was obvious in hindsight.