Enterprise AI Architect — I make operational context move automatically across chat, tickets, wiki and CRM.
The hard part is mapping the same people, teams, customers, projects and decisions across systems that use different vocabularies. A temporal knowledge graph resolves that mapping and shares each accepted fact where it belongs. Delegated reads, human-approved writes and audit trails keep the flow safe enough for production.
New: merged into tetherto/qvac — the official repo of Tether's QVAC AI platform: #3729 · validate RAG queries before logging
(Aug 2026) · all contributions ↓
Map once, share continuously. The context layer connects each system's vocabulary so accepted facts reach the right people and tools; the execution and assurance layers govern what moves and prove what happened.
- In production: operational context layer: maps people, teams, customers, projects and decisions across chat, tickets, wiki and CRM, then automatically shares accepted facts where they belong. Reads run as the signed-in person, never a service account; writes wait for a human decision on every door, including MCP. Case study with synthetic screens
- agentic-task-system
— CLI + MCP that turns the systems you already run (TickTick, Notion, GitHub, Obsidian, Airtable, Google Tasks) into agent context. Hybrid RRF retrieval, provenance on every result, reviewed writes, undo. The real systems stay the source of truth.
- graphiti-local
— local-first temporal knowledge graph: six read-only MCP retrieval tools, a
kgCLI, human-gated fact ingestion. Built on Graphiti, runs on your own hardware. Answers "what was true at the time?", not just "what is true now?"
- draftcat
— governed AI pipelines for service businesses: deterministic-first, operator-approved, single Go binary (MIT). Tool-call gates, HITL, audit trail.
- agent-approval-gate — the
draft → validate → approve → dispatch → auditpattern with JSON schemas, an n8n workflow, and an email-approval example.
- processbench
— business-specific benchmark packs for repeatable workflow regression checks. Swap a model, measure what actually moved.
- skillgate
— deterministic finish-line gates for AI coding agents: a model-independent evaluator that blocks commit/publish until your definition-of-done passes. opencode, Claude Code, pre-commit, CI.
Enterprise integration: Azure, Microsoft Teams, Jira, Confluence, HubSpot, LeanIX — from delivery work in regulated environments, not from demos.
Production AI Automation Notes — a running set of repos and gists on:
- permission-aware retrieval that never leaks through counts, titles or provenance chains (#19)
- knowledge graphs people fill through verdicts, never through an agent (#18)
- approval gates for AI agents
- MCP server security (permissions, tenants, audit logs)
- Claude Code policies for company repos
- n8n workflows with human approval
- audit-log schemas
Follow if you're building agents that need to work outside demos.
| Project | PR | What |
|---|---|---|
| tetherto/qvac | #3729  |
Validate RAG queries before logging them |
| Tencent/WeKnora | #2614 |
GPT-5 / o-series vision model support (max_completion_tokens) |
| Tencent/WeKnora | #835 |
Parallel tool calling support |
| steveyegge/beads | #2884 |
Multi-project support, Notion sync, backup/restore |
| e2b-dev/infra | #2273 |
Local dev docs: prerequisites, verification steps, troubleshooting |
| pacifio/cersei | #10 |
Native Google Gemini provider + Cohere & SambaNova support |
A separate track from the enterprise work — same engineering standard, different audience.
- capcut-cli
— independent CLI to edit CapCut / JianYing projects: subtitles, timing, speed, templates, cut long-form → shorts. No API; reads
draft_content.jsondirectly. - browserground — local UI-grounding specialist for hybrid AI agents: Qwen3-VL-2B LoRA, screenshot + target → strict JSON bbox. Drop-in for Claude Code, Codex, browser-use.
Stack: Python, Go, TypeScript, Node.js, Kubernetes, Linux, systemd, vector databases, temporal knowledge graphs, MCP, LLM APIs.
Website: renezander.com · YouTube: @determa
A gist series on production-grade patterns for AI pipelines. Each entry covers one pattern with copy-pasteable code.
| # | Topic | Pattern |
|---|---|---|
| #1 | Agent Approval Gates | Schema-validated proposals, human review, deterministic dispatch |
| #2 | Token Budgets | Per-step, per-pipeline, per-day enforcement |
| #3 | Agentic Knowledge Base | Karpathy-style LLM wiki with pluggable adapters |
| #4 | CapCut / JianYing from an LLM agent | Deterministic JSON command boundary |
| #5 | SQLite Dedup + Crash Safety | WAL mode, seen_items, audit log |
| #6 | Prompt-Injection Defense | Input sanitization, schema validation |
| #7 | PDF Cite Verification | Auditable LLM extraction with per-fragment bounding boxes |
| #8 | Stateless JSONL Queue Runner | One job per line in, one result per line out; no daemon, no open port |
| #10 | Deterministic Step Pipelines | Fixed typed steps; the LLM never picks the next action |
| #11 | Pipeline Fixture Testing | Dry-run pipelines from JSON fixtures; zero API calls, deterministic CI |
| #12 | LLM Skills as YAML | Prompt + output_schema + role in versioned YAML; validated by a linter |
| #13 | Inbound Agent Webhook Auth | Constant-time bearer token, fail-closed on empty secret, async 202 dispatch |
| #14 | Self-Improving Voice Agent | Human-approved prompt diffs; harvest > group > propose > 2 approval gates > git commit > staging smoke > auto-versioned publish |
| #15 | AI Action Audit Trail | Append-only action_approvals table: who approved which payload, when; gate-violation query; GDPR Art. 22 provenance |
| #16 | Agent Memory That Degrades | Keyword fallback when the embedding provider dies; the read path never fails outright |
| #17 | Validate Before You Log | RAG query guard above the first log line, plus the ordering test |
| #18 | Never Let an Agent Write to the Knowledge Graph | Queued proposals, one writer; a refusal that exits 0 is not a refusal |
| #19 | Dropped Without a Count | Delegated reads, membership at question time; cite what the caller may open or drop it |
Reference implementation for entries #1, #2, #5, #6, #7, #9, #10, #11, #12, #13: draftcat (Go, MIT).