SupportOps AI
A multi-tenant support platform that grounds responses in authorized company knowledge and keeps consequential AI actions under human control.
- My role
- Full-stack AI engineer · Product architecture, Python APIs, retrieval, agent workflows, evaluation, security, and deployment
- Technology
- Next.js
- Python
- FastAPI
- PostgreSQL
- pgvector
- Redis
- Docker
- LLM APIs
01
The problem
Support teams lose time searching fragmented documentation and repeatedly reconstructing answers, while a conventional chatbot can introduce unsupported claims or act beyond a user's authority. SupportOps AI needed to make institutional knowledge useful inside the ticket workflow without hiding evidence, crossing tenant boundaries, or allowing the model to perform consequential actions without review.
02
Architecture and decisions
I kept Next.js focused on the product interface and placed the authoritative backend in FastAPI. PostgreSQL owns organizations, tickets, conversations, permissions, and evaluation records, while pgvector supports tenant-scoped semantic retrieval. The retrieval pipeline combines keyword and vector candidates, reranks them, and requires cited evidence before an answer is returned. Agent tools expose a narrow set of support operations, and every state-changing action passes through an explicit approval step. Redis-backed workers isolate document ingestion and evaluation runs from request handling, while structured traces capture retrieval, model, tool, latency, and cost data.
03
Key features
- Multi-tenant ticket, conversation, knowledge-base, and role-based access workflows
- Asynchronous document parsing, chunking, embedding, indexing, and failure recovery
- Hybrid retrieval with reranking, tenant filters, grounded answers, and source citations
- Support tools for response drafting, classification, escalation, and ticket updates
- Human approval before state-changing agent actions
- Repeatable evaluations for retrieval, citations, grounding, tool selection, latency, and cost
04
Outcomes
- 01
Combined 2 retrieval channels—keyword and semantic candidates—before reranking evidence for each response.
- 02
Constrained 4 AI capabilities—drafting, classification, escalation, and ticket updates—behind explicit operational boundaries.
- 03
Tracked 6 evaluation dimensions while enforcing 3 safety gates: authorization, tenant-scoped retrieval, and human approval.
Metrics are included only where they were measured or documented.