Get to know the Human ME
Currently at Kop - Building cool stuff
Love learning cool stuff, and design things that actually matters
Ifrane, MA
AI Platform Engineer
--:--:-- · IFRANE
my name is

Youness

Youness Aboukad · systems
Open to platform & agent work

I build the layerwhere agents stop being demos

Youness Aboukad
the operator
second monitor

I’m Younessan AI platform engineer in Ifranewho cares about the queue that backs up and the state that drifts

Agent OrchestrationReal-time SystemsLLMOpsDesignGamefy lifeContext and Harness Engineering

Business analytics student with a passion for AI, automation and prodcut design. Blending in technical curiosity with entrepreneurial drive, building tools that turn ideass into practical systems, ambitious, love fun, and focused on solving problems that drive impace

Systems in production

Four things that carry real traffic. Each one started as a diagram and ended as something on call.

System 01
2025 — now

Real-time AI Interview Agent

An autonomous agent runs live technical interviews — holding a WebRTC session open, executing the candidate’s code mid-conversation, and defending its own prompt against injection.

LiveKitRedisJudge0PostgreSQL
Read the log
candidate · webrtccode · judge0state · redisrecord · postgresAGENTturn 04SESSION OPENPROMPT GUARD ONmodular monolith · real-time path never waits on the slow one
TRACEsession.trace
System 02
2025 — now

AI Prompt Evaluation Platform

Prompts stop being guesses once you can measure them. Fetch → Audit → Optimize → Validate scores every revision against a regression gate, so a change that reads better but performs worse never reaches production.

FastAPIStrapiPython
Read the log
prompt revision · job description · interview setupFETCHAUDITOPTIMIZEVALIDATEREGRESSION GATEreads better but scores worse → never ships
PYgate.py
System 03
2024 — now

Ksawyoux

A multi-agent assistant that lives in Telegram. Routine work answers on a sub-second FastLLM stream; anything that needs to think forks off to parallel heavy agents and rejoins.

CrewAILiteLLMpgvectorPython
Read the log
messagefast streamsub-secondresearchreasonactparallel heavy agentsmergedtwo-speed routing · depth without the wait
PYrouter.py
System 04
2024

Lume

The conscious wallet — a fintech engine correlating physical recovery data with spending behaviour, on the bet that money decisions are physiological before they are rational.

90% coverage60% faster infraBiometrics
Read the log
recovery signalthresholdspend decisionsHELDHELD90% test coverage · 60% faster infrastructuremoney decisions are physiological before they are rational
SQLsignal.sql

The track so far

Every role here ended with something automated that used to be done by hand.

The bench

What is actually open on the second monitor.

Claude CodePythonSQLFastAPIPostgreSQLRedisNext.jsn8nFigmaJiraGeminiChatGPT

Bachelor in Business Intelligence · Al Akhawayn University in Ifrane

BI Professional CertificateGoogle
Associate Cloud EngineerGoogle
AI Enterprise WorkflowIBM
Game TheoryStanford Online

Engineering logs

9 write-ups on the systems above, what broke, what the fix cost, and what I would build differently.

01

Project 1: Architecting the AI Eval System

A comprehensive platform for AI prompt optimization, job description generation, and automated interview setup.

10 min read
02

Project 2: Engineering the Ghost in the Machine: A High-Concurrency AI Interview Architect

Architecting a modular monolith for live AI interviews with real-time code evaluation.

15 min read
03

Project 3: Archgraph, or How to Teach a Codebase to Answer Questions About Itself

A six-component Go workspace that fuses Git history, ASTs and ownership metadata into a living property graph you can interrogate in plain English.

14 min read
04

Project 4: Kraft, or Letting an AI Do Real Work Without Handing It the Keys

A planning and artifact system that turns a goal into an inspectable plan, runs approved AI craft work, keeps every version, and refuses to let prompt text approve anything.

16 min read
05

From Data to Emotion: Building a Biometric-Aware Fintech Engine, the Future of Financial Wellness

Architecting a high-concurrency engine correlating real-time biometrics and transactional data to solve the emotional blind spot in finance.

12 min read
06

Organizational Refactoring: Engineering the Human Middleware at AIESEC

Transforming legacy manual environments into data-driven automated machines by treating organizational processes like technical pipelines.

7 min read
07

The Human API: Architecting Requirements at Atlasia

Eliminating Business-Technical Debt by acting as the translation layer between high-level vision and granular execution.

8 min read
08

The Quality Quest: Bridging Blueprints to Deployment

Developing comprehensive blueprints and rigorous testing frameworks.

4 min read
09

How I Built a Multi-Agent AI Assistant in Telegram

From a simple stateless chatbot to a proactive, multi-agent personal assistant living in my messaging app.

8 min read

Let’s talk

I’m most useful on the problems where a clever demo has to survive real users — the transport that has to stay open, the state that has to stay honest, the prompt that has to stay measured.

Comment
Youness

Open to platform and agent work, full-time roles, and long arguments about where the state should live.

Contact
Youness Aboukad