AI Systems
We build language-model features and agent systems that behave like software: grounded in your data, observable, and designed for the moments when the model gets it wrong.
Problem → Approach
Getting a model to produce an impressive answer is easy. Getting it to produce a reliable one inside a real product is the work. We treat the model as one component — surrounded by retrieval, memory, validation and a fallback path — and build the risky parts first so you know early whether the idea holds up.
What we build
- LLM-driven features, such as the generated dialogue in MEMORABLE
- Multi-agent systems that coordinate through shared memory, as in Syntapse
- Knowledge-graph memory and retrieval with Cognee and Memgraph
- Hosted models through Groq, or open models run locally through Ollama
What this covers
AI agents and multi-agent systems
Agents earn their place when a task has several distinct steps that each need their own context. In Syntapse, five specialised agents — ideation, product requirements, system architecture, stack evaluation and implementation planning — coordinate through one shared knowledge graph instead of a prompt chain, so changing one decision reaches every agent that depends on it. We design agent systems around explicit shared state like this, so their behaviour can be inspected and corrected rather than guessed at.
Retrieval and knowledge graphs (RAG)
A language model only knows what it is given. Retrieval connects it to sources you control, so answers are grounded in your data instead of the model's memory. Syntapse uses a Cognee knowledge graph on Memgraph as its memory layer; each agent pulls the context it needs from the graph before it calls the model, and writes its results back.
AI chatbots and website assistants
The assistant on this website is one of ours. It runs on Groq, answers only from verified information about ALLIET, keeps its API key on the server, validates every request, and ignores attempts to rewrite its instructions. We build assistants the same way for products and company websites: scoped to what they should know, and honest when they don't know something.
Hosted and local models
We choose where a model runs based on latency, cost and where your data is allowed to go: fast hosted inference through Groq, as in MEMORABLE's generated dialogue and this site's assistant, or open models run locally through Ollama when data should stay on your own machines, as in Syntapse.
Technologies we use here
Related work
Further reading: Shared memory instead of prompt chains: how Syntapse coordinates AI agents
Questions
Can you add an AI assistant to an existing website or product?
Yes. The assistant on this site is a small server-side API route that calls the model, with its knowledge, rate limits and input validation kept on the server, so it can usually sit alongside an existing site rather than requiring a rebuild.
Do you work with only one model provider?
No. Our work so far uses Groq-hosted models and open models run locally through Ollama. The right choice depends on latency, cost and where your data is allowed to go.
What happens when the model gets something wrong?
We design for it: grounded context, validated output, and a fallback path. MEMORABLE, for example, switches to hand-written dialogue whenever the model is unavailable, so the game keeps working.