I design and operate end-to-end autonomous AI systems — from zero-to-one architecture through production. My work sits at the intersection of multi-agent orchestration, fail-closed safety, and LLM evaluation. The systems I build run with zero human oversight at enterprise scale.
Current system in production (TNG Shopper, 2024 → present):
- 11 enterprise clients across 5 countries (ES · PT · IL · US · MX) — ~10.5M product pages under autonomous management at $0.0006 / page
- 7-node multi-agent directed acyclic graph with ~73.5M agent operations per full run (weekly executions + daily price-delta runs) · 234 managed websites
- Gemma 4 26B-A4B Mixture-of-Experts on self-hosted vLLM with PagedAttention inference. Multi-Low-Rank-Adaptation research documented in the forensic runbook.
- Originality, Relevance, Accuracy, Value — four-axis multi-model evaluation with fail-closed policy at 68.9% pass rate by design · Deterministic Evaluation and Monitoring Audit System enforcing every boundary
Ten years to get here. Six projects. The pattern: how to unblock human-dependencies. See the profile time-spine →.
graph LR
subgraph Inference["Inference Layer"]
MoE["Gemma 4 26B MoE<br/>vLLM · PagedAttention"]
L1["LoRA α"]
L2["LoRA β"]
L3["LoRA γ"]
MoE --> L1 & L2 & L3
end
subgraph DAG["7-Node Autonomous Pipeline"]
N1["City DNA<br/><sub>Context</sub>"] --> N2["Normalizer<br/><sub>4 sub-agents</sub>"]
N2 --> N3["Synonyms<br/><sub>Expand</sub>"]
N3 --> N4["SV Gate<br/><sub>Filter</sub>"]
N4 --> N5["Writer<br/><sub>Generate</sub>"]
N5 --> N6["Validator<br/><sub>O-R-A-V</sub>"]
N6 --> N7["Features<br/><sub>Vectorize</sub>"]
end
subgraph Eval["Evaluation & Safety"]
ORAV["O-R-A-V Judge<br/><sub>Multi-Model Scoring</sub>"]
DEMAS["DEMAS Audit<br/><sub>JIT · Fail-Closed</sub>"]
end
L1 & L2 & L3 --> N1
N6 --> ORAV
DEMAS -.->|"intercept at<br/>every boundary"| N1 & N2 & N3 & N4 & N5 & N6 & N7
ORAV -.->|"RL feedback<br/>prompt mutation"| N5
style MoE fill:#1a1a2e,stroke:#0A84FF,color:#fff
style ORAV fill:#1a1a2e,stroke:#30D158,color:#fff
style DEMAS fill:#1a1a2e,stroke:#FFD60A,color:#fff
Node Anatomy — Each node contains multiple sub-agents
Every directed-acyclic-graph node is a bounded ecosystem, not a single LLM call:
| Layer | Role | Example |
|---|---|---|
| Deterministic Gate | Schema validation, type coercion, regex | Pydantic, Python AST |
| Probabilistic Agent | Semantic extraction, classification | Gemini Vision, SLM |
| Autonomy Layer | Originality-Relevance-Accuracy-Value scoring, confidence thresholds | Multi-model consensus |
| Memory | Long-term state, prompt cache mutation | Redis LTM, Firestore |
The deterministic gate always fires first. The LLM is invoked only if the gate passes.
AI & ML
Infrastructure & MLOps
Evaluation & Safety
| Year | Project | What it proved |
|---|---|---|
| 2016 | Asset (Sept 2016 — 2019) | Three years of solo contractor work for new-stage startups: web setups, ERP→web ETL by hand, spreadsheet automation, business plans. The data-transformation reps every later pipeline compounded on. |
| 2019 | Data Mining (Feb 2019 — Jul 2020) | Five-stage manually-orchestrated pipeline for an Israeli financial-services firm. ₪50M+ in new Assets Under Management. A pipeline is a series of filters, not a series of steps. |
| 2020 | Seller App (Jan 2020 — Apr 2024) | Computer vision for retail digitization. 3 computer-vision generations · 60M+ canonical Stock-Keeping Units · $10K Monthly Recurring Revenue plateau. The architectural origin of retrieval-grounded computer vision. |
| 2020 | Tasko AI (Oct 2020 — Dec 2023) | Production agentic system before the term existed. 21,102 labeled tasks · 153 clients · 1,561 intent patterns · 4-layer Classify / Retrieve / Execute / Verify. |
| 2024 | Elysium (2024 — 2025, paused-pending-Pipeline) | Physical-Context AI for Retail. 13 brands validated · 15,600+ store locations · 15 patent claims (3 independent + 12 dependent). |
| 2024 | Pipeline Observatory (2024 — present) | The synthesis. Seven-node directed acyclic graph, deterministic gates first, fail-closed by default. 10.5M product detail pages / month · 73.5M ops / month · $0.0006 / page · 68.9% pass rate across Originality, Relevance, Accuracy, Value. |
| Repository | Description |
|---|---|
| agent-dag-pipeline | Open-source distillation of the seven-node directed acyclic graph. Google Agent Development Kit + Vertex AI + four-axis Originality-Relevance-Accuracy-Value evaluation + Direct Preference Optimization data flywheel. |
Antigravity-OS · pip install ag-os |
Orchestration infra that lets a small team ship more product per engineer, safely. Built at TNG (2-engineer dev team) to streamline the SDLC by reducing human-dependencies in the dev loop: AI dev assistants run inside a 9-rule constitutional policy-as-code, with Cost Guard, Flight Recorder for replay, Self-Healing CI, Dreaming Module for offline agent self-improvement, drop-in MCP server. |
| gemma4-vllm-deployment | Forensic runbook documenting 20 failure modes across 30+ deployment versions of Gemma 4 Mixture-of-Experts on Vertex AI with vLLM. The community reference for production Mixture-of-Experts serving. |
| pipeline-observatory | Source of the live observability site at manzela.github.io/pipeline-observatory. Architecture visualization — Mixture-of-Experts sparse routing, causal directed-acyclic-graph tracing, live execution telemetry. |
| WP-Multisite | A WordPress multisite for retailer sites, written from scratch to be natively understood by AI search bots (ChatGPT, Claude, Perplexity, Google AI Overview). Explicit AI-bot ALLOW contracts (OAI-SearchBot, ChatGPT-User, PerplexityBot, ClaudeBot, Applebot, Amazonbot, DuckAssistBot), llms.txt indexing surface, JSON-LD with SpeakableSpecification for voice. Sanitised from the TNG production stack (Sage 10 + custom sunrise.php + 1,056-LOC schema generator). |