/projects/elderai
A customer-facing sales operations analyst for sales leaders: routes questions to fast or analysis agents, queries a read-only Postgres sample database, delegates to specialist subagents, runs sandboxed Python analysis, and persists dashboards, charts, PDFs, CSVs, tables, and image artifacts.
Sales leaders need trustworthy answers from call data, transcripts, coaching signals, and competitive intelligence, but generic chat over a database can leak raw SQL, fabricate metrics, or generate disposable reports.
Elder AI routes requests into fast or analysis agents, uses read-only Postgres tools for evidence, delegates bounded work to data/coaching/artifact subagents, runs sandboxed Python only when needed, and persists dashboards, charts, tables, PDFs, CSVs, and images.
The product must feel like a customer-facing sales operations analyst while coordinating auth, RLS, conversations, tool events, artifact storage, local versus Vercel sandbox modes, and deterministic product-taste evals.
The system treats artifacts as durable conversation objects. Instead of dumping analysis into chat, it creates reopenable manager-facing assets backed by verified rows, scoped tools, and policy-aware response modes.
This demonstrates Stefan's strength in business-agent product architecture: RevOps domain modeling, read-only data access, artifact UX, AI SDK ToolLoopAgent patterns, Supabase persistence, sandbox execution, and eval-backed response quality.
Elder AI is a serious business-agent case study because it connects chat UX to governed data access, specialist subagents, sandboxed analysis, durable artifacts, Supabase persistence, and product-taste evals.
The app separates fast lookup from deeper analysis before the agent starts using tools. That keeps simple manager questions responsive while reserving sandbox, artifact, and subagent work for requests that need it.
The database path is read-only and domain-specific. Calls, transcripts, coaching skills, metadata, and competitive-intel extractions are queried through constrained tools instead of handing the model raw database authority.
Tables, charts, dashboards, PDFs, CSVs, images, and dataset analyses are durable objects, not transient chat decorations. They can be reopened, discussed, and persisted alongside conversations.
The architecture distinguishes weak local Python workspace isolation from production-shaped Vercel Sandbox execution. That matters because generated analysis code should not quietly inherit application secrets.
Keeps chat UI, API routes, RSC demo, artifact pages, and deployment in one application boundary.
Keeps routing, artifacts, tool schemas, and agent handoffs explicit across a large product surface.
Provides ToolLoopAgent, streaming, RSC streamUI, and model-provider integration.
Powers the customer-facing analyst, routing, image generation, and specialist agents.
Handles auth, RLS persistence, conversations, artifacts, and organization membership.
Acts as the sales-call source of truth with read-only query access for the agent.
Separates fast answers from analysis and artifact workflows.
Retrieve manager evidence without giving the model write-capable database access.
Runs derived metrics, calculations, and generated files in a scoped workspace.
Delegates bounded metric work instead of sending the entire conversation to another agent.
Focuses on transcript snippets, skills, objections, and coaching risks.
Turns verified rows and metrics into dashboards, charts, PDFs, tables, and exports.
Saves results so users can reopen and discuss prior analysis.
Checks whether responses stay manager-facing and avoid internal tool leakage.