Yumi Nexus
Founder & AI engineer · self-directed, pre-revenueMar 2026–PresentA production-deployed, multi-tenant generative-AI platform that turns natural-language intent into secure, observable action. Built and operated solo in three layers with strictly one-way dependencies: no layer sees the one above it, and the core ships as the versioned package the platform installs.
- Agent loopA streaming turn loop over provider-native function calling, with persistent memory across turns. Three hosted providers — Claude, Gemini, OpenAI — plus local Ollama and fastembed behind one interface, sharing error classification and a warm-up / pull / shutdown lifecycle.
- Tool exposureThree registration modes, chosen so the agent gathers before it speaks: autorun runs ahead of generation and injects only its result; pinned stays in the list; dynamic is retrieved against the query.
- PromptsLayered prompt composition under a semantic version plus a content hash, so wording edited without a version bump is caught automatically.
- MemoryTwo stores by design: SQLite as the source of truth, LanceDB for vector search, reconciled both ways; during an index rebuild concurrent writes commit to SQLite only, so they cannot duplicate rows.
- ShippingFunction-calling bindings for 11 targets, a CLI and a turn inspector; 699 tests with mocked providers, CI on Python 3.10–3.13, released to PyPI by OIDC Trusted Publishing.
L2
Multi-tenant platform
yumi-enterprise
closed source
closed source
- TenancyPer-tenant isolation over PostgreSQL, with Argon2 hashing, OAuth, linked chat accounts and auth rate limiting.
- MeteringQuotas measured in turns, tokens and dollars across 21 admin endpoints and a React console.
- Observability & monitoringA console for finding faults: an aggregate view of how requests were routed to tools, and the exact execution trace behind any single turn.
- Governance & privacyTwo tiers enforced in SQL, not policy: history views select only message length, never content, and revealing a turn writes the audit record first, failing 503 if that write fails.
- DeploymentGCP Compute Engine behind Caddy/TLS with Docker Compose, health checks, traces and metrics; serves yumi.nexus, the identity site and a live agent demo.
- IdentityOne Firebase-backed sign-in serves the site, the agent and every edge app, with cross-app handoff; authorisation lives in hand-written Firestore rules.
- Infrastructure as codeThe production GCP project is managed in Terraform on a locked remote backend — the running estate imported and held to a clean “No changes” plan first, then five reviewed changes, including SSH moved to IAP-only.
- Continuous deliveryA push deploys itself over GitHub OIDC into Workload Identity Federation — no stored cloud credential: three repos rebuild and health-check behind a host lock, any failure rolls all three back, and the public hostname is then re-verified.
- ContinuityScripted backups with retention, and a restore drill actually executed: snapshot to a fresh disk, restored onto a network-isolated VM, then 21 tables diffed.