Ten years, six systems shipped. Today: a pipeline serving 11 enterprise clients across 5 countries — every step has a safety check before the AI runs.
Active 2016 → 2026TNG Shopper · current5 prior shipped systems
Now · 2024 → Present
What's running right now
product pages under autonomous management
11 clients
5 countries (ES · PT · IL · US · MX)
234 websites
$0.0006 / page
I lead the autonomous content pipeline at TNG Shopper — seven steps, each one a safety check before the AI runs. 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. Outputs are scored on four things — Originality, Relevance, Accuracy, Value — and only 68.9% pass. The rest get rejected.
The work · 2016 → 2026
Six projects. The pattern: how to unblock human-dependencies.
Each one taught me something I still use today. The earliest taught the reps; the rest compounded.
2016
Junior Years
Three years of freelance work for startups and small companies — web development, inventory imports, spreadsheet automation, market research.
·3 startups, 2 low-techs, side-projects · Solo
Three years of contractor work nobody else wanted — the data-transformation reps that every later pipeline compounded on.
6 categories of work
3 startups, 2 low-techs, side-projects
ETL by hand · ERP → website
LessonDoing data transformation, normalization, enrichment, and formatting by hand for three years builds the reflex that makes every later pipeline feel like the same job at scale.
BridgeThe same extract → normalize → enrich → format pattern, scaled up: Data Mining.
Automated lead-finding pipeline for an Israeli wealth-management firm.
·Israeli Financial Services Firm · Solo
A pipeline is a series of filters, not a series of steps.
17 mo in production
₪50M+ in new Assets Under Management
5-stage manually orchestrated pipeline
LessonThe five-stage filter is the architectural primitive everything later compounds. Frame any system as a chain of deterministic filters before adding probabilistic compute.
BridgeEleven months in, a parallel co-founded venture began — a camera-first retail computer-vision tool: Seller App.
Camera-first onboarding tool that turned retail stockrooms into online storefronts.
·Co-founded with Dr. Eli Osherovich · Bootstrapped 4 yrs
Three computer-vision generations — barcode, then image, then video — against a canonical catalog of 60 million Stock-Keeping Units. Bootstrapped four years.
Physical-context AI platform for retail demand prediction — 13 brands signed before pause.
·Physical-Context AI for Retail · Paused-pending-Pipeline
A platform fusing search intent, geo-coordinates, and foot-traffic data into store-level demand signals — eleven specialized agents, fifteen patent claims.
13 brands validated
15,600+ store locations
15 patent claims (3 independent + 12 dependent)
11 specialized agents
LessonThe right product can be the wrong order of operations. Pause is a routing decision, not a failure mode.
BridgeSequenced behind a prerequisite layer: the Pipeline ships first; Elysium re-activates downstream once it's operating across enough tenants to feed it.
Vertex AI · self-hosted Gemma 4 26B-A4B Mixture-of-Experts on vLLM 0.17.2rc1 · PagedAttention · Google Cloud Storage FUSE · NVIDIA A100 80GB
Orchestration & data
Google Agent Development Kit · Python · FastAPI · Pydantic · Jinja2 · BigQuery · Google Cloud Storage · Firestore · Redis · Cloud Functions
Observability & governance
Langfuse · OpenTelemetry · Model Context Protocol · Deterministic Evaluation and Monitoring Audit System for fail-closed safety · Four-axis evaluation across Originality, Relevance, Accuracy, Value · Direct Preference Optimization via Reinforcement Learning from AI Feedback
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.
Governance kernel for AI agents. Open Policy Agent-style policy-as-code, Model Context Protocol server, cost guard, "Dreaming Module" self-improvement loop.
