AI systems
Production agents and workflows — not demo theater.
Humoud Al Munawer
Founder · Operator
I build, test and explain
technology systems.
- AI Systems
- Automation
- Platforms
- Venture Building
Global perspective. Based in Kuwait.
Research what works, share judgment, and ship systems others can run.
02
Domains
What I work on
Four lanes. One loop: validate → build → explain → scale what earns it.
Automation
Repeatable systems that cut grind, keep judgment.
Platforms
Durable product surfaces others can rely on.
Venture building
Focused ventures with a reason to exist.
03
Guides · tools · frameworks
Latest useful work
Reviewed guides that should outlive a social feed. Mechanisms, numbers, and when to use them.
Guide
Prompt caching: cut LLM input cost up to 90%
One line of configuration turns repeated context from full price into a tenth of the cost.
GuideBatch API: cut LLM costs 50% when work can wait
Route non-urgent calls through Batch instead of the standard endpoint — same call, half the price.
ResearchOpen-vs-closed AI gap, measured
UK AISI cyber evals on GLM-5.2 and DeepSeek-V4-Pro — where the gap is months, and where it still holds.
04
Public objects
Builds & experiments
Verified public tools only — Problem / Built / Result / Learning.
Founder Agent Skills OS
- Problem
- Agent sessions fail founders when tip-spam replaces a thin constitution and clear jobs.
- Built
- Personal operator pack: thin AGENTS.md plus five founder skills — research, decide, scan rivals, stress a spec, close the day.
- Result
- Drop-in markdown for Claude Code, Cursor, Codex and similar — no vendor lock-in, no marketplace required.
- Learning
- Operators need jobs with kill criteria more than another “invent an OS” prompt.
Model Routing Calculator
- Problem
- Teams default every task to the top-tier model and overpay for judgment that does not need it.
- Built
- A dependency-free Python CLI (`route.py`) that recommends model tier by complexity and estimates cost vs. always-top-tier.
- Result
- Example outputs: ~63% of baseline for judgment-heavy work; ~46% for long-unattended tasks.
- Learning
- Routing is a product decision, not a model beauty contest — measure before you standardize.
05
Editorial
Insights
Longer judgment — security, agents, and operating notes worth keeping.
Featured
Research
Framework
Framework
Library
Founder note
The open-vs-closed AI gap, measured in cyber evals
Where months separate open and closed models — and where the gap still holds on harder tasks.
Cost beats effort when you route agent work
Benchmark takeaway: choose models by job shape, not brand loyalty.
Public notes from LinkedIn, X, and Instagram
Sourced posts on MCP OAuth, prompt injection, automation, and shipping loops — collected in one place.
06
Where work becomes companies
Ventures
Humoud is the founder-operator. VentraM and FANN are the ventures. Commercial execution → VentraM. Independent product → FANN.
VentraM
Validate → Build → Scale. For qualified commercial systems work, go to VentraM.
Visit VentraM →
FANN
Live site (founding access): verified marketplace for physical art — provenance-first records, certificate-backed ownership, resale royalties. Apply for the founding cohort at tryfann.com.
Visit FANN →07
Founder-operator
About Humoud
I build and test technology systems, then explain what holds up. Kuwait is home base; the work is global. The GCC is an important market and perspective — not a brand boundary.
This site is the founder-operator layer: public judgment, guides, and tools. Ventures sit under that layer — VentraM for commercial execution, FANN as an independent product.
Hierarchy: Humoud (founder) → Ventures (VentraM, FANN).
08
Follow the work
Connect
Social channels for judgment and public work. WhatsApp for founder notes. VentraM for qualified commercial execution.
Commercial door
Have something worth building?
Qualified commercial systems work — AI, automation, platforms, venture building — goes to VentraM. Validate → Build → Scale.