Founder & CEO, TNG Shopper · Senior AI Product Manager
Autonomous multi-agent AI systems in production, at enterprise scale — no humans in the loop.
**10.5M** product pages under autonomous management · **234** websites · 11 enterprise clients · 5 countries · **$0.0006** marginal cost per page · **68.9%** quality-gate pass rate (fail-closed by design)
figures re-rendered nightly from production telemetry · per-client evidence: Google Search Console case studies · last audit: 2026-07-21
| System | Scale, verified | Live proof | |
|---|---|---|---|
 |
TNG Shopper — autonomous content pipeline: 7-node gated DAG, self-hosted Gemma 4 26B MoE on vLLM | 10.5M pages · 234 sites · 11 enterprise clients · 5 countries | Pipeline Observatory · GSC evidence |
 |
Atelier — autonomous design agent, Google for Startups AI Agents Challenge 2026 | Deterministic gates before any LLM judge · multi-judge consensus · DPO self-improvement flywheel | atelier.autonomous-agent.dev · Repo |
 |
agent-dag-pipeline — the open-source fail-closed agent DAG behind the pipeline | ~33,000 production LLM calls across 4 model generations | Live telemetry · Video |
 |
ag-os — governance kernel for AI agents (pip install ag-os) |
  |
PyPI · Repo |
Status dots are re-probed nightly. A system that stops answering shows a hollow "unverified" dot; a persistently retired endpoint is demoted to "archived" — a live dot is never shown on a dead link.
Selected engineering. Every claim maps to committed code — clone and verify, or read the auditor's notes in the appendix below, where each load-bearing figure is re-derived weekly in CI.
| Repository | One verified outcome | Signal |
|---|---|---|
| Atelier | Deterministic gates always precede probabilistic agents; hermetic, offline make verify; per-axis provenance on 5 rubric judges |
Google AI Agents Challenge 2026 |
| agent-dag-pipeline | Fail-closed 7-node DAG: O·R·A·V multi-model evaluation, immutable failure records, DPO + LoRA self-improvement flywheel |  |
| Antigravity-OS | Zero-LLM governance kernel: policy-as-code constitution, solvency gate, flight recorder, full audit trail |  |
| gemma4-vllm-deployment | Every failure mode on the road to Gemma 4 MoE on vLLM / 1×A100-80GB, documented with its fix — plus the single working config | Forensic runbook |
| WP-Multisite | The production publishing surface behind the pipeline, sanitized verbatim — 1,056-line JSON-LD generator, AI-search-aware llms.txt / robots.txt |
Production blueprint |
The secret isn't making the AI smarter — it's building a strict system of rules that automatically catches and drops the AI's mistakes.
| Layer | Mechanism | Committed in |
|---|---|---|
| Spend | Per-user lifetime token cap enforced server-side before any model call | Atelier |
| Spend | Daily hard cap at the LLM proxy · alerts at 50 / 75 / 90 / 100% | AutonomousAgent |
| State | Immutable FailureRecord on every exit · store-context integrity asserted at ENTRY and PRE_EXPORT |
agent-dag-pipeline |
| Policy | Policy-as-code constitution · solvency gate · retry limit with human escalation · zero-LLM core | Antigravity-OS |
| Input | Prompt-injection screening wraps every model call | Atelier · agent-dag-pipeline |
| Training | RL kill-switches on cost overrun and eval regression · promotion only on measured eval gain | AutonomousAgent |
Operating principles
- Deterministic gates in front of probabilistic agents — never the reverse. Invalid input is rejected before a single token is spent.
- Fail closed. Every pipeline exit is a recorded, causally-traced failure; budget caps are enforced server-side before any model call.
- Self-improvement is governed. DPO/RLAIF flywheels promote adapters only past gated evals, with kill-switches on cost overrun and regression.
|
Recent releases |
Recent writing
|
Google Cloud Generative AI Leader (active) · Google Cloud Professional ML Engineer (Jun 2026) · Product Experts Certified Product Manager (active) · CrewAI Multi-Agent Systems (active) · 32 Google Cloud credentials · Guest lecturer, Bar-Ilan University (May 2025)
Named inventor on 15 patent claims in physical-context AI — an 11-agent system fusing search-intent and geospatial data to forecast street-level demand across 15,600+ store locations. The Elysium case study
Building TNG Shopper full-time. Talking to enterprise retail teams, partners, and investors about autonomous content operations.
danielq1603@gmail.com · LinkedIn — I read every message.
Colophon — this page maintains itself: a nightly GitHub Action re-renders every figure and health probe, and a weekly audit job re-derives every load-bearing claim from a fresh clone of its source repo. Anything it cannot verify is stamped stale — never invented. The nightly workflow · The weekly audit · The build script
Appendix — stack, auditor's notes & the longer story
Where I'm fluent
agents google-adk · vertex-agent-engine · mcp · a2a · litellm · crewai
training dpo / preference-tuning · lora / s-lora / qlora · rlaif flywheels · ppo / grpo
serving vllm (moe, multi-lora, a100) · cloud-run · gke · fastapi · gcsfuse
evals llm-as-judge (o·r·a·v) · multi-judge consensus · golden trajectories · axe / lighthouse
guardrails model-armor · policy-as-code (ag-os) · budget caps · tiered sandboxes · opa · vault
telemetry opentelemetry · cloud-trace · langfuse · arize phoenix · bigquery
data apache beam / dataflow · pub/sub · postgres · redis · vector dbs · uber h3
platforms gcp (vertex ai · bigquery · gke) · kubernetes · terraform · wordpress multisite · shopify
languages python · go · node.js · typescript · react
Auditor's notes — every load-bearing number on this page is re-derived from a fresh clone, weekly, in CI. A figure that can't be reproduced is stamped stale, not repeated.
| Claim | Value | Derived by | Verified |
|---|---|---|---|
| Test functions (Atelier) | 1,315 in 120 files | grep -rcE '^\s*(async )?def test_' --include='*.py' on fresh clone |
2026-07-20 |
| Test functions (AutonomousAgent) | 2,011 in 198 files | grep -rcE '^\s*(async )?def test_' --include='*.py' on fresh clone |
2026-07-20 |
| Test functions (Antigravity-OS) | 133 in 9 files | grep -rcE '^\s*(async )?def test_' --include='*.py' on fresh clone |
2026-07-20 |
| Test functions (agent-dag-pipeline) | 61 in 4 files | grep -rcE '^\s*(async )?def test_' --include='*.py' on fresh clone |
2026-07-20 |
| O·R·A·V gate thresholds (agent-dag-pipeline) | 0.6 / 0.7 / 0.8 / 0.6 | constants in agent_dag/config.py, fresh clone |
2026-07-20 |
| Constitution rules (Antigravity-OS) | 9 | count of templates/rules/*.md on fresh clone |
2026-07-20 |
| Documented failure modes, forensic runbook (gemma4-vllm-deployment) | 19 | numbered ### sections in docs/FORENSIC_RUNBOOK.md, fresh clone |
2026-07-20 |
Also in the lab — AutonomousAgent: production-hardened self-improving agent running 24/7 on GCP — 5-tier sandboxing, OpenTelemetry tracing, fail-closed failure matrix, MoE-routed RL self-training
Before TNG Shopper — CTO & AI Product Engineer at Tasko AI: a WhatsApp-native assistant for 153 paying clients resolving 1,561 distinct user intents, pre-trained on 21M+ messages · a camera-first retail onboarding tool (0 → $10K MRR, 3 computer-vision pipelines, 60M+ SKU catalog) · a 6-stage lead-qualification pipeline that drove $17M+ in new assets under management
The longer story (2013 → now) — Elysium · Tasko AI · Seller App · Data Mining · Junior Years · Essays
Numerals on this page are set in Fraunces (OFL), pre-outlined so they render identically everywhere. English · Hebrew · Russian.