Agentic AI with an audit spine, multi-tenant Postgres, generative UI. One engineer. One paying design partner. No PM, no four-week sprint. This is what AI-native product delivery looks like in 2026 when you skip the framework tax and write the audit layer yourself.
A multi-tenant supplier-reconciliation product for an accountancy running on Xero. The end user opens a chat surface and asks which suppliers can I pay today or why is this one in variance. The LLM calls typed tools that return generative-UI components — tables, PDF viewers, drill-downs, action buttons. Every tool call writes an immutable audit row. Every PDF view is logged. Every reconciliation override is signed by a human.
30% feature. 70% audit / RLS / tenant isolation / observability / "who did what and when." That ratio is what lets the Big-4 audit partner sign off. Most AI demos skip it. This one does not.
Problem. Per-tool audit logging leaves holes. The auditor's first question always lands on the one call that wasn't logged.
Fix. Wrap audit at the registry, not per-tool. One choke point.
const tool = createAuditedTool({
name: 'mark_reconciled',
execute: async (args, ctx) => { ... },
})
// every tool shipped through this factory is audit-wrapped at the boundaryProblem. Multi-tenant Postgres RLS silently leaks when an edge function under service-role forgets to set tenant_id.
Fix. The DB derives tenant_id from the parent and rejects orphans. RLS policies stay simple.
CREATE TRIGGER set_tenant_id BEFORE INSERT ON statement_lines
FOR EACH ROW EXECUTE FUNCTION derive_tenant_from_statement();Problem. Chat tool, UI, and CLI doing the same soft-delete via three paths means three different bugs.
Fix. One RPC. Tenant-scoped. Re-entrant. Fully auditable. Every caller takes the same path.
CREATE FUNCTION delete_statement_line(p_line_id uuid, p_reason text)
RETURNS jsonb LANGUAGE plpgsql SECURITY DEFINER AS $$
-- soft-delete, override matches, log, return state
$$; client signal → diagnose → ship → audit log proves it
↑ │
└─────────────────────────────────────────┘No PM. No four-week sprint. The engineer is the PM, diagnosis happens in the audit log plus the LLM transcript plus the canary output, and the ship cadence is the meeting cadence.
Send me your repo and a bottleneck you're trying to get through. I diagnose and scope a 5-day proof in 48 hours. You pay nothing until you see the proof working.
mordechaipotash@gmail.com →I parachute in, build the audit spine + tool registry + first 5 production tools, hand off a pattern your team extends.
10–20 hrs / week. I own the agentic AI surface end-to-end while your team owns the rest of the product.
For staff / principal AI engineer full-time roles — open to talk.
"Design and implement core protocol features. Build SDKs. Drive open-source technical discussions on agent infrastructure. Deep familiarity with multi-tenant systems, observability, and audit primitives."
— representative 2026 staff AI engineer JD
Not 1% of applicants have shipped this combination in production this year. This week is the proof.