The stack we use to ship enterprise AI agents.
We're vendor-agnostic by design — but we have strong opinions about what good looks like. Here's the technology we use to build production AI agents for enterprise customers, and the stacks we integrate with.
Cloud platforms
We deploy on whichever hyperscaler our customers already trust. Our platform is portable across all three.
- Amazon Web Services (Bedrock, Lambda, S3, RDS, SES)
- Microsoft Azure (Azure OpenAI, Functions, Blob, AKS)
- Google Cloud Platform (Vertex AI, Cloud Run, BigQuery)
- Customer VPCs / on-prem (BYOC deployments)
AI models
Model-agnostic by design. We pick the best-fit model per task — not per vendor.
- Anthropic Claude (Opus, Sonnet, Haiku) — our most-used
- OpenAI GPT (GPT-4 Turbo, Codex) — for mechanical tasks
- Google Gemini (Pro, Flash) — for formatting and lighter loads
- Open-weight models (Llama, Mistral, Qwen) — for on-prem / sensitive data
- Customer-hosted endpoints (any OpenAI-compatible API)
Enterprise systems we integrate with
Our agents plug into the systems of record customers already run. Our deepest expertise is SAP — the rest we add as customers bring them.
- SAP (S/4HANA, ECC, EWM, TM, BTP, ABAP) — deepest expertise
- Oracle ERP Cloud / E-Business Suite
- Microsoft Dynamics 365 (F&O, Business Central)
- NetSuite, Workday
- Salesforce, ServiceNow, HubSpot
- Custom in-house ERPs (REST, SOAP, JDBC, MCP)
Data & analytics platforms
Where customer data lives — and where our agents read from / write to.
- Databricks (Delta Lake, Unity Catalog)
- Snowflake
- Google BigQuery, AWS Redshift, Azure Synapse
- PostgreSQL with pgvector (our default for our own services)
- Vector stores: pgvector, Pinecone, Weaviate (situational)
Languages & frameworks
A small, sharp tech surface. Picked for productivity, type safety, and long-term maintainability.
- TypeScript (Next.js 15, Node 20+, React)
- Python (FastAPI, async-first)
- Go (where latency matters)
- Prisma + PostgreSQL for data modelling
- TailwindCSS for design systems
Integration patterns
How we connect AI agents to enterprise systems without breaking either side.
- Model Context Protocol (MCP) — first-class for tool calls
- Webhooks + event-driven architecture
- Direct REST + SOAP for legacy ERPs
- SFTP / file-drop for batch reconciliation
- CDC (Debezium, Fivetran) for warehouse sync
Security & compliance
Enterprise-grade from day one. Customers in regulated industries have specific requirements — we meet them.
- SSO via Okta, Azure AD, SAML, OIDC
- TLS 1.3 in transit, AES-256 at rest
- Per-tool, per-agent RBAC
- Immutable audit trails on every action
- Data residency: EU (default), UK, India, US — your choice
- SIG Lite, DPA available on request
DevOps & observability
The boring infrastructure that makes the difference between a demo and a product.
- Docker, GitHub Actions, Kubernetes
- PM2 / systemd for service supervision
- Datadog, New Relic, OpenTelemetry
- Sentry / GlitchTip for error tracking
- Postgres dumps + S3 backups, daily
How we choose technology.
Boring is a feature
We choose proven technology over novelty. Postgres over the trendy database. Next.js over the framework-of-the-month.
Vendor-agnostic
No allegiance to any one cloud, model, or ERP. Customer outcome first, vendor preference last.
Type-safe by default
TypeScript end-to-end where it matters. Prisma for the data layer. Schema validation at every boundary.
Observable from day one
Every action logged. Every reasoning step captured. Replay incidents end-to-end.
Small surface area
A small, sharp set of tools. Easier to maintain, harder to break, faster to onboard new engineers into.
Security as a first-class concern
We don't bolt on auth, encryption, or audit trails after the fact. They are part of the original design.
Running something we don't list?
Our integration layer reaches new systems in days, not months. Tell us what you run and we'll tell you what's involved.
See our security page for full detail on encryption, residency, and audit trails.