No mystery middleware, no vendor lock-in theater. This is what actually runs when you ask Squinty a question.
what happens when I hit send?
Your request is routed by meaning, planned by the Architect lane, turned into SQL by the Analyst lane, run read-only against your database, rendered as widgets, and traced end to end. About two seconds, most days.
Languages and frameworks
Python 3.12
The API, planners, background workers, and ETL
Starlette + Uvicorn
The HTTP layer under every endpoint
Svelte 5 + SvelteKit
The entire interface, from Studio to generated apps
TypeScript
Type-safe frontend services and widget contracts
Vite
Frontend build tooling
Data and storage
PostgreSQL 15
The platform database, with row-level security doing the isolation
Qdrant
Vector search over documents, schemas, and prompt routing
Redis
Caching and coordination
NATS JetStream
Messaging backbone for ingest and background jobs
S3-compatible storage
Durable storage for uploaded files and artifacts
CloudNativePG
Runs PostgreSQL inside the cluster
AI and machine learning
OpenAI, Anthropic, xAI
Cloud reasoning lanes, chosen per user
Ollama, LM Studio, LocalAI
Local model runtimes for on-premises and air-gapped setups
Prophet
Time-series forecasting behind anomaly alarms
NumPy + pandas
Statistical alarm evaluation and data processing
ONNX
Portable format for registered prediction models
Documents and content
pypdf
PDF text extraction, page by page
python-docx
Word documents, headings and tables included
openpyxl
Excel sheets and cells
OCR and vision providers
Turn images into searchable, citable text
Frontend
Chart.js
Every chart type, with zoom and interactive legends
GridStack
The drag-and-resize dashboard grid
driver.js
The guided first-run tour
Web Speech API
Voice input in the composer
Security and identity
OpenID Connect
Single sign-on via an Auth0 broker
TOTP
Authenticator-app two-factor
Envelope encryption
Per-user, per-record keys for API keys and memories
JWT
Programmatic API access
Observability
OpenTelemetry
Every request, query, and model call in one trace
Grafana Tempo
Trace storage and visualization
Journey tracking
Full request lifecycles for support and debugging
Infrastructure and quality
Docker + Compose
Single-machine deployments
Kubernetes + Helm
Cluster deployments, k3s included
Nginx
The gateway
pytest + golden corpora
Every prompt change replayed against known-good evaluations
Ghost Inspector
Daily end-to-end browser tests against live staging
Questions about the architecture?
Happy to go deep: lanes, tenancy, tracing, local model sizing. Bring your security team, we like them.
