01Selected builds
The systems we build, described as engineering.
Catalyst Labs is a new studio, and we would rather show you our engineering than dress up placeholder case studies. Everything below is a reference architecture — our own work, labelled as exactly that.
- Published
- No client case studies yet
- Architectures
- 4 documented
- Invented claims
- None
/Reference architectures
Open any build to see how it is put together.
4 documented
Problem
Enquiries land in four inboxes. Whoever sees one first answers it, the rest go cold, and nobody can say how many arrived last month.
What changes
No enquiry depends on a person noticing it. Response time becomes a property of the system rather than of who was online.
Technology
- Next.js
- Node.js
- PostgreSQL
- LLM
- WhatsApp API
Architecture
- Capture from web form, WhatsApp and ad channels into one queue
- Model qualifies and tags each enquiry against your own criteria
- Routing to the right owner with an SLA timer
- Automated first response and follow-up sequence
- Every enquiry written to the CRM with its full trail
Reference architecture · not a client case study
Problem
The real process lives in a shared sheet, three chat groups and one person's memory. Status is whatever the last message said.
What changes
Status stops being a question someone has to ask. The process is written down in software, so it survives staff changes.
Technology
- Next.js
- TypeScript
- PostgreSQL
- Redis
- Cloud
Architecture
- Model the actual workflow — including the exception paths
- One record per job, with state, owner and history
- Role-based views: floor, supervisor, management
- Alerts on stalls, breaches and anything stuck
- Reporting built on the same records, not a re-export
Reference architecture · not a client case study
Problem
Documents arrive as PDFs and photos, and a person retypes the numbers into a system, one at a time, with the error rate that implies.
What changes
Keystrokes are replaced by exception handling. People only touch the documents the system is genuinely unsure about.
Technology
- Python
- FastAPI
- Vector DB
- LLM
- PostgreSQL
Architecture
- Ingest from email, upload and shared drives
- Extraction into a strict schema with per-field confidence
- Validation against existing records and business rules
- Review queue for anything below the confidence threshold
- Write-through to the accounting or ERP system
Reference architecture · not a client case study
Problem
Most incoming questions have documented answers, but the documentation isn't where the customer is, so the team answers them again.
What changes
Answered-before questions stop reaching a human, and the ones that do arrive already summarised with the history attached.
Technology
- Next.js
- Vector DB
- RAG
- LLM
- Webhooks
Architecture
- Index the real knowledge: docs, past tickets, policies, product data
- Retrieval-grounded answers with citations back to source
- Confident answers auto-resolve; the rest escalate with context attached
- Gaps in the knowledge base surfaced from failed answers
- Deployed to the channels customers already use
Reference architecture · not a client case study
Want one of these built against your process instead of a generic one? That conversation is the first stage of every project.
Start a ProjectWant one of these against your process? Not a generic one.
Every build starts by mapping how the work actually moves through your business. The architecture follows from that, not from a template.