E R A Y

Backend systems · RAG · Human-in-the-loop

SupportFlow AI

A learning-focused support copilot for Turkish B2B SaaS teams—designed to retrieve company knowledge, prepare cited answer drafts, and leave every customer-facing decision with a support agent.

← Back to Projects
FastAPITyped REST API and consistent error contracts
Tenant-awareVerified organization context and scoped access
Async-readyRedis-backed Celery worker foundation
ReproducibleDocker Compose, migrations, tests, and CI

The problem

Support teams lose time searching product documentation and previous resolutions. Generic chatbots add another risk: confident answers without evidence.

SupportFlow is built around a stricter contract—retrieve relevant company knowledge, show the sources behind a draft, return insufficient context when evidence is weak, and require a person to approve the result.

Product principles

  • Sources before confidence. Claims must be traceable to retrieved material.
  • No answer beats invention. Insufficient context is a valid result.
  • People keep control. AI output is never sent automatically in v1.
  • Isolation is mandatory. Every organization sees only its own data.

Modular monolith

System direction

supportflow / request lifecycle

Implemented FastAPI boundary Authentication, validation, error contracts
Implemented Tenant context Organization membership and scoped tickets
Implemented Data & jobs PostgreSQL, Alembic, Redis, Celery
Next Retrieval & citations pgvector, document chunks, evidence
Next Agent decision Approve, edit, reject, audit

Implemented foundation

  • FastAPI application factory with liveness and versioned API routes.
  • Argon2id password hashing, short-lived access tokens, and database-verified memberships.
  • Organization-scoped ticket and message models with service-layer status transitions.
  • PostgreSQL schema managed through reversible Alembic migrations, with pgvector-ready local infrastructure.
  • Redis/Celery worker path, Docker Compose environment, and automated quality checks.

Next product layers

  1. Document workflowSecure uploads, extraction, versioning, and background processing.
  2. ClassificationProvider-neutral adapters and validated structured output.
  3. Grounded RAGTenant-filtered retrieval, citations, and no-answer thresholds.
  4. Human approvalApprove, edit, reject, feedback, and complete audit events.

My contribution

  • Led the PostgreSQL, Redis, Docker Compose, testing, and initial CI foundation.
  • Contributed across the shared backend, authentication, ticket, documentation, and integration workflow.
  • Helped define development standards that keep migrations, tenant boundaries, and operational checks explicit.

SupportFlow AI is a collaborative project by Emir and Eray. Feature leadership rotates so both contributors can design, implement, review, and explain the complete system. The shared source repository is currently under private development.

Engineering stack

Python 3.13 FastAPI Pydantic PostgreSQL SQLAlchemy Alembic Redis Celery Docker pytest Ruff Pyright GitHub Actions