1. Home
  2. AI Development Services
  3. AI Technology Stack
Engineering choices

AI Technology Stack

We are deliberately model-agnostic and cloud-flexible. What follows is what we build with and, more usefully, how we decide between the options.

Engagements from $10,000 to $100,000+Serving USA · UAE · UK · Canada · EuropeYou own the code and the IP
The problem

How we choose, and why it is not a vendor decision

Every component below is replaceable. That is the point — the architecture is layered so a model or database swap is a configuration change, not a rewrite.

Model selection is benchmarked against a sample of your real tasks, not against public leaderboards. A model that leads a general reasoning benchmark can lose badly on your specific extraction task, and the difference in cost per thousand requests at production volume is often larger than the difference in quality.

The same applies to vector databases. For most workloads under a few million chunks, PostgreSQL with pgvector is the correct answer and adds no new operational surface. Dedicated vector databases earn their place at scale, or where you need specific hybrid-search and filtering behaviour.

  • Benchmarked against your data before selection
  • Model provider abstracted behind a service layer
  • Two providers configured for failover on critical paths
  • Open-weight models where residency requires it
  • Cost per transaction modelled at projected volume
  • No component chosen because of a partner incentive
Architecture

Reference architecture

The shape most of our production systems take, regardless of use case.

01 InterfaceReact / Next.js app, in-product widget, WhatsApp, telephony or headless API.
02 API layerASP.NET Core or FastAPI with authentication, validation and rate limiting.
03 OrchestrationLangChain / LangGraph or a custom state machine for agent control flow.
04 RetrievalEmbeddings plus hybrid search over pgvector, Pinecone, Qdrant, Weaviate or Azure AI Search.
05 Model layerProvider-abstracted calls with routing, caching, retries and failover.
06 Tools & dataTyped connectors to CRM, ERP, databases and internal APIs.
07 ObservabilityTracing, evaluation runs, cost tracking and drift alerts.
Technology

The AI technology stack we build on

We are deliberately model-agnostic. Selection is driven by your accuracy requirements, data-residency constraints and cost per transaction — not by a vendor relationship.

Large Language Models

OpenAIAnthropic ClaudeGoogle GeminiMeta LlamaMistralAzure OpenAI Service

AI Frameworks

LangChainLangGraphHugging FaceLlamaIndexSemantic Kernel

Vector & Search

PineconeQdrantWeaviatePostgreSQL + pgvectorAzure AI SearchElasticsearchRedis

Backend

ASP.NET CorePython (FastAPI)Node.jsCelery / queuesgRPC

Frontend

ReactNext.jsTypeScriptTailwind CSS

Data

PostgreSQLMongoDBSQL ServerSnowflakeKafka

Cloud

Amazon Web ServicesMicrosoft AzureGoogle Cloud

Infrastructure & MLOps

DockerKubernetesTerraformGitHub Actions CI/CDMLflowLangSmith / OpenTelemetry
Disclosure: Technologies are listed as tools we build with. Ezulix does not claim certified partner status with any model or cloud provider on this page. [VERIFY: partner/certification list before publishing]
Security & Governance

Security-conscious architecture, from the first design review

Enterprise AI fails on governance more often than on models. Every system we build is designed to support enterprise security requirements and to give your risk team answers rather than assurances.

Data privacy & residency

Your data stays in the region and tenancy you nominate. We architect for no-training-on-your-data configurations and document exactly which vendor endpoints see which fields.

Role-based access control

Retrieval and tool permissions inherit your existing roles. A user cannot surface a document through the AI that they could not open directly.

Authentication & authorization

SSO via OIDC/SAML, short-lived tokens for agent tool calls, and per-tool scopes so an agent holds the narrowest possible privilege.

Encryption

TLS in transit, AES-256 at rest, managed keys via your cloud KMS, and encrypted vector stores for embedded content.

API security

Gateway-level authentication, signed webhooks, IP allowlisting, request validation and quota enforcement on every exposed endpoint.

Audit logging

Every prompt, retrieval, tool call, model version and human override is logged with a trace ID, so any output can be reconstructed months later.

Data isolation

Per-tenant separation at the storage, index and key level for multi-entity groups and regulated environments.

Secure prompt handling

System instructions are server-side, user content is treated as untrusted input, and we test against prompt-injection and tool-abuse patterns.

PII protection

Detection, masking or tokenisation of personal data before it reaches a model, with configurable redaction policies per field.

Human approval workflows

High-impact actions — payments, refunds, contract sends, record deletion — route to a named approver instead of executing autonomously.

Monitoring & anomaly detection

Alerting on unusual tool usage, cost spikes, refusal rates and quality regressions.

Rate limiting & abuse control

Per-user and per-tenant throttles, spend caps and circuit breakers so a runaway loop cannot become a runaway invoice.

Secure deployment

Private networking, secrets in a managed vault, immutable builds, dependency scanning, and infrastructure as code.

On compliance: Ezulix designs compliance-ready architecture aligned to frameworks such as GDPR, HIPAA and SOC 2 control objectives. Certification status for any specific standard should be confirmed directly with our team before contract. [VERIFY: current Ezulix certifications]
FAQ

Questions enterprise buyers ask us first

Which LLM should we use?
It depends on the task, and usually more than one. Complex reasoning and long-document work tends to favour the larger frontier models; high-volume classification and extraction often runs well on a smaller, cheaper model at a fraction of the cost. We benchmark candidates against a sample of your real inputs during architecture and route different tasks to different models where it saves money without costing accuracy.
Which vector database do you recommend?
PostgreSQL with pgvector for most projects — it is fast enough well into the millions of chunks, and it means no additional system to operate, back up and secure. Pinecone, Qdrant or Weaviate when you need scale beyond that or specific filtering and hybrid-search behaviour. Azure AI Search when you are already committed to Azure and want managed hybrid retrieval with minimal assembly.
Do you build in .NET or Python?
Both, and often together. Python for the AI and data layer where the ecosystem is strongest; ASP.NET Core or Node.js for the application and API layer, particularly where it matches your existing estate. We will match your team's stack where you plan to maintain the system in-house.
Is Ezulix a certified partner of these providers?
This page lists technologies we build with. Any partner or certification status should be confirmed directly with our team before it forms part of a procurement decision — we do not make partnership claims in marketing copy. [VERIFY: current partner status]
Project brief

Talk to an AI solution architect

No junior sales rep, no discovery deck. The person on the call is the person who will design the system.

  • Response within one business day
  • Mutual NDA signed before detailed discussion
  • Written scope, one price, one delivery date
  • You own all source code, models and IP at launch
Email: sales@ezulix.com [VERIFY]

We use these details only to prepare your scope and estimate. Your idea stays yours — mutual NDA before any detailed discussion.

Next step

Bring us the process that is costing you the most.

Book a 45-minute call with a solution architect. You leave with a use-case shortlist, a reference architecture sketch and a realistic build envelope — whether or not you build it with Ezulix.