Freedom is no longer a dream.
It’s an engineering problem.

Now · 2024 → Present

What's running right now

• 5 countries (ES · PT · IL · US · MX)
• 234 websites
• $0.0006 / page

I lead the autonomous content pipeline at TNG Shopper — seven steps, each gated before the model executes. The model is Google's Gemma 4 (a 26-billion-parameter Mixture-of-Experts model), self-hosted on our own infrastructure. Every full run does about 73.5 million sub-tasks.

Also building

1. 01

   Atelier

   Experimental · Google for Startups AI Agents Challenge 2026

   An autonomous design agent that converges instead of one-shotting. On Google ADK and Vertex AI (Gemini), it generates a UI candidate, rejects it against fail-closed accessibility and design-token gates, scores it across several axes, then fixes and retries until it clears a quality threshold. A per-user token budget — hard cap, circuit breaker, fail-loud — keeps cost bounded. Each run logs scored trajectories toward a Vertex tuning job; the full retrain-and-promote loop isn’t closed yet. Experimental, with a live demo up.

   See it running →
2. 02

   Autonomous Agent

   Experimental · Phase 1 (local)

   A production-hardening layer around Nous Research’s Hermes agent — the operability scaffolding a long-running, tool-using agent needs to run safely: inline secret-scrubbing on the model proxy, a codified failure-mode taxonomy with handlers, a budget watchdog, 24-hour-silence escalation, a network-isolated tool sandbox, and one-command halt-and-snapshot. Models route to Vertex AI through a LiteLLM proxy. The self-improving control plane — model routing and RL self-training — is designed and partly built, not yet wired; Phase 1 runs locally today.

   View on GitHub →

The work

Six Products

Flawless automation is actually born from deep human expertise, capturing the exact memory, context, and iterative learning of an expert, and then scaling it infinitely with machines.

1. 01

   TNG Shopper

   Live · 2024 → today

   I built an autonomous content pipeline that helps multi-location retail brands automatically generate SEO-optimized product pages without human involvement. By combining deterministic rules with a highly efficient AI (Mixture-of-Experts) that learns from its own mistakes, the system reliably produces millions of localized web pages a month at a fraction of a cent per page.

2. 02

   Elysium (Part of TNG Shopper)

   Paused on purpose · 2024 → 25

   I built an AI chat inference that helps retail stores understand local customer searches, location data, and engagement. 13 major brands have confirmed they need this technology. We paused Elysium, because it needs a large amount of data to make accurate predictions. Right now, TNG Pipeline is doing the heavy lifting to collect that data. Once we have enough, we will turn Elysium back on.

   Read the case study →
3. 03

   Tasko AI

   Production · 2020 → 23

   I built a WhatsApp AI assistant for an Israeli services company, trained on 21 million customer messages. We had an advanced AI completing complex tasks years before it became a tech trend, but the iterative feedback loop still required manual adjustments and continuous evaluations. I realized how significant it is to fully automate this process. To perfectly replicate human execution, the AI requires automated reinforcement learning that dynamically adjusts context, memory, and personalized preferences for every unique user and task combination.

   Read the case study →
4. 04

   Seller App (1st Startup)

   Bootstrapped 4 years · 2020 → 24

   I built an AI-powered app that lets physical retailers instantly turn their store shelves into an online shop using just their phone’s camera. Instead of manually typing out product details, users simply point their camera to scan items, and the app automatically builds and publishes a digital catalog.

   Read the case study →
5. 05

   Data Mining

   Sole engineer · 2019 → 20

   I built a custom data-mining pipeline for an Israeli wealth management firm to generate highly targeted sales leads, as existing tools like ZoomInfo didn’t work for the local market. By extracting data from local websites, enriching and ranking prospects based on wealth and geography, I delivered leads that generated over ₪50M in new client portfolios for Eitam Finance (the Wealth Management firm).

   Read the case study →
6. 06

   Junior Years

   Warehouse to web · 2013 → 19

   I started working at age 14, spending years in physically demanding jobs like construction, factory manufacturing, and warehousing. While working as a personal assistant, I began teaching myself web development and digital automation to solve problems for clients. Those early years of exhausting manual labor taught me a critical lesson: trading time for money isn’t scalable, which fueled my lifelong drive to build software that infinitely scales my impact.

   See the early reps →

Key takeaways

What each experience taught me

1. The Pipeline Observatory — Real-world AI needs a strict safety net.

   If you want AI to run 24/7 without a human watching it, 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 before they ever reach production.

   Source: Pipeline (2024 — present).

2. Elysium (Physical-xA-Intelligence) — You can’t skip the data-gathering phase.

   Predictive AI sounds incredibly exciting, but it is completely useless without a massive foundation of reliable data. You have to build the unglamorous data pipelines first before the “smart” AI can actually work.

   Source: Elysium (2024 — 25, paused on purpose).

3. The WhatsApp AI Assistant — Context is everything.

   Making an AI “smart” isn’t just about using a bigger model. To truly automate complex tasks, the AI needs to dynamically remember the unique context, history, and preferences of every single user it interacts with.

   Source: Tasko AI (2020 — 23).

4. The Data Mining Pipeline — Niche problems beat generic tools.

   Giant, expensive software platforms often fail in local markets. By building a custom, highly localized tool specifically for Hebrew-speaking clients, I was able to generate millions in value that the global tools completely missed.

   Source: Data Mining (2019 — 20).

5. The Seller App — Don’t rely on users to do the heavy lifting.

   Our tech worked perfectly, but the business hit a wall because we relied on busy store employees to manually scan products. True automation has to run in the background without asking humans to change their daily habits.

   Source: Seller App (2020 — 24).

6. My Junior Years (Manual Labor) — Trading time for money has a hard ceiling.

   Working exhausting physical jobs showed me that there are only so many hours in a day. The only way to create massive, global impact is to build technology that scales infinitely.

   Source: Junior Years (2013 — 19).

Stack · Across ten years

Where I’m fluent

I own the entire product lifecycle—from cloud infrastructure to the user experience. Below is the cross-section of tools I use to ship autonomous software.

Languages & IaC

Python · TypeScript · Go · Shell · HCL

ML & Deep Learning

PyTorch · Transformers (BERT, Attention) · Multimodal RAG · TensorFlow

LLMs & GenAI Models

Gemma 4 26B MoE · vLLM · NVIDIA A100

Agentic Orchestration

Vertex AI Agent Engine · Google Agent Development Kit (Multi-Agent Architectures) · Model Context Protocol · LiteLLM · LangChain · CrewAI · AG2 (AutoGen) · SOPS

AI Product Strategy

B2B Enterprise SaaS · Go-to-Market (GTM) Strategy · Telemetry-Driven Iteration · ROI & Pricing Modeling · O-R-A-V / HEART Evaluation

Product Analytics & BI

Mixpanel · Looker Studio · GA4 · BigQuery ML (BQML) · Monday.com (Custom Workflows)

GCP Ecosystem & AI

Vertex AI (Model Garden, Vector Search, Gemini Enterprise, Agent Builder) · Dataflow (Apache Beam) · Pub/Sub · Cloud Run

Distributed Systems & Cloud

Kubernetes (GKE) · Temporal · Modal · Docker · Terraform · HashiCorp Vault · Cloudflare · GitHub Actions · Debian

Backend & Data Layer

FastAPI · Pydantic · PostgreSQL · Redis · Next.js · Uber H3 Geo-Indexing · BrightData · WebScraper · FireCrawl · TinyFish · Honcho · Pinecone · Vector Search · SQLite · Chroma

MLOps & AI Governance

Vertex AI MLOps · Model Armor · Langfuse · OpenTelemetry · Open Policy Agent (OPA) · Arize Phoenix · Healthchecks

Web Architecture

React · Bun · Vite · Vercel · WordPress / Sage 10 (Auto-scaling Multisite) · PWA · Wix · Shopify

Quality & Testing

pytest · mypy · Ruff · Playwright

Code · Personal projects + OSS

Code you can read

• Antigravity-OS pip install ag-os

  Open-source governance kernel for autonomous NHITL AI systems. Open Policy Agent-style policy-as-code, Model Context Protocol server, cost guard, “Dreaming Module” self-improvement loop.

• Autonomous Agent

  An operability layer around Nous Research’s Hermes agent: inline secret-scrubbing on the model proxy, a failure-mode taxonomy with handlers, a budget watchdog, 24h-silence escalation, a network-isolated sandbox, and halt-and-snapshot. Routes to Vertex AI via LiteLLM. The self-improving control plane is designed and partly built, not yet wired.

• agent-dag-pipeline

  Open-source distillation of the 7-node DAG behind the TNG Shopper pipeline. Google ADK + Vertex AI + ORAV evaluation + DPO data flywheel.

• gemma4-vllm-deployment

  Forensic runbook of 20 deployment failure modes for Gemma 4 MoE on Vertex AI + vLLM. The doc I wish I’d had.

Open-source contributions most-recent

• anthropics / claude-code issue #59401
• Agent-Field / agentfield PR #573 · review comment
• NousResearch / hermes-agent issue #26383
• And more — see github.com/Manzela for the full ledger.

Recent issues · Essays

Recent essays

• The Physical-Context Flywheel Kairi-AI · Nov 2025
• Epiphany Medium · Nov 2025

Earlier writing

• The Marco Economy of Hybrid Retail Medium · Nov 2022

And more — see medium.com/@manzela for all essays.