system-prompts-and-models-of-ai-tools/dealix/docs/AI_STACK_DECISIONS.md

AI Stack Decisions — Dealix v3

Current immediate stack

• FastAPI — primary API surface.
• Supabase / Postgres / pgvector — project + strategic memory (see migration + setup doc).
• Deterministic internal agent runtime — SafeAgentRuntime patterns in v3 routers; no silent outbound execution.
• Deterministic Personal Operator — Arabic briefs, opportunities, drafts with approval_required.
• Sentry / OpenTelemetry — optional instrumentation in api/main.py when dealix.observability is present.
• Project Intelligence (local) — scan_project, chunk_text, naive_search, CLI scripts/index_project_memory.py.

Next stack (when credentials and policy are ready)

• Langfuse — tracing, evals, prompt versioning (dependency already in project).
• OpenAI Agents SDK — structured traces, tool calls, guardrails.
• LangGraph — durable workflows with human-in-the-loop approvals.
• LlamaIndex — RAG over project_chunks + strategic_memory.
• WhatsApp Cloud API — interactive reply buttons (max 3 per message).
• Gmail drafts API — outbound only after approval.
• Google Calendar — schedule drafts first; create events only after explicit approval.

Later stack (cost / scale driven)

• vLLM / Ollama — when inference cost or latency justifies self-hosting.
• n8n / MCP — after internal policy and audit trails are stable.
• Dify — prototyping only; not the core production execution path.

Why this shape

• Dealix owns memory schema, PDPL posture, approval contracts, and audit semantics.
• External frameworks are replaceable execution engines around the same memory and policy core.