The right stack
for the task.
Available models, your business context, and software that fits the workflow. The final architecture follows the requirements of each project.
Models & Providers
Choose a model for the task, data constraints, response quality, latency, and operating cost. These are representative options to evaluate for a project.
OpenAI APIs
Provider APIs for integrating language model capabilities into custom applications.
Claude
Anthropic models available through provider APIs for application and agent workflows.
Gemini
Google models available through provider APIs for custom AI applications.
Open models
Consider managed or self-hosted models when the task and deployment constraints support them.
Knowledge & Retrieval
Connect approved information to the application. Retrieval design includes ingestion, access permissions, source citations, and checks for missing evidence.
PostgreSQL
Relational storage for application records, workflow state, and structured information.
pgvector
Vector similarity search within PostgreSQL for retrieval workflows.
Document ingestion
Parse, organize, and refresh source material before making it available to retrieval.
Source permissions
Apply the same access boundaries to retrieved information as the underlying documents.
Agents & Tool Integration
Give an agent a defined job and explicit tool boundaries. Keep consequential actions behind validation and appropriate human approval.
Tool interfaces
Expose narrowly scoped operations with clear inputs, permissions, and error handling.
Structured outputs
Validate model responses against the data shape required by the application.
Workflow state
Track progress, retries, and handoffs across steps instead of relying on a single prompt.
Approval queues
Let a person review proposed actions before the system changes important records.
AI Automation
Connect AI steps with existing software through APIs and workflow tools. Choose the integration approach around the systems already in use.
Make
A visual workflow tool for connecting applications and coordinating automation steps.
Zapier
An application automation platform for connecting events and actions across tools.
Webhooks & APIs
Event-driven connections for receiving updates and invoking approved operations.
Python & TypeScript
Application code for custom logic, validation, integrations, and AI interfaces.
Evaluation & Operations
Make quality visible before and after release. Agree on what to measure and how the team should respond when a workflow needs attention.
Task evaluation sets
Representative inputs and expected behavior for comparing prompts, models, and system changes.
Retrieval checks
Review evidence relevance, source coverage, and whether generated answers are supported.
Trace monitoring
Inspect tool calls, failures, latency, and usage to understand system behavior.
Operating playbooks
Document escalation, recovery, access management, and ongoing review responsibilities.
Architecture shaped by requirements.
Model choice, data access, integrations, and evaluation criteria form a connected engineering process.