Epistemic Noise
I build systems that learn from the world.I write about why we struggle to.
I'm Aissam Outchakoucht, an AI engineer working across language, speech, and intelligent systems. I use this space to share things I build, ideas I'm exploring, and essays about uncertainty, perspective, intelligence, and the strange business of knowing things.
The rooms
Four places to look
A personal site, not a brochure. Pick a room.
01
About
AI engineer working on conversational systems, and the questions they leave open.
Open →
02
Projects
Darija, voice, chatbots, and production AI systems, from early shipping to current work.
Open →
03
Notes
Short observations on building, evaluating, and operating AI systems.
Open →
04
Essays
Longer writing on certainty, perception, and how much of what we call knowing actually stands.
Open →
Work
Things that shipped
Darija, voice pipelines, chatbots, and production AI systems.
The Darija TTS Model Hunt: Failures, False Starts, and a Working Voice
Three text-to-speech experiments, 20.73 hours of synthetic speech, and a working OmniVoice model: lessons in model selection, evaluation, and training on rented GPUs.
What I Would Teach Differently About Machine Learning Today
I reopened an ML/DL course I designed and taught at EMSI. Most of the mathematics survived. My idea of where the hard engineering lives changed considerably.
Getting More Work Out of One H100
Lessons from reducing a multi-model production AI batch pipeline's runtime by 75% on a single H100, covering profiling, I/O, batching, CPU/GPU overlap, model scheduling, caching, backpressure, precision and resilient batch execution.
From Speaker Matching to System Design: Building a Production-Scale Voice Pipeline
What started as a small experiment matching voices across a handful of calls became a local pipeline that backfills tens of thousands of recordings and runs daily, where the hard part stopped being the model and became the system.
How We Expanded DODa to 500K Darija Examples
Darija is inconvenient for NLP: two scripts, flexible spelling, heavy borrowing. DODa’s latest expansion turns that inconvenience into a controlled 500K-row training corpus.
Teaching TranslateGemma Moroccan Darija: What Worked, What Broke, and What 32.8 BLEU Actually Means
Fine-tuning TranslateGemma 4B for Arabic-script Darija and Arabizi, moving the run between GPU clouds late in training, and figuring out whether 32.82 BLEU was actually good.
From Request to Architecture
Start AI system design from the guarantees a product must preserve: access, actions, evidence, and survival of interrupted work.
Darija libraries
Python libraries for Darija ⇆ English: word distance, translation lookup, and an LLM-assisted translator that does not pretend Darija is just messy Arabic.
Alexa Skills with Python : An Introduction
An Introduction to Alexa Skills
Darija Open Dataset
The largest open collaborative dataset for Moroccan Darija ⇆ English, built for NLP, now past 500,000 semantic rows, in both Arabic and Latin script.
WhatsApp and Messenger chatbots
Business chatbots on WhatsApp and Messenger that talk to a real backend: Dialogflow, Twilio, Flask, Firebase.
Writing
Recent pages
Notes, essays, and projects: the latest across the rooms.
Occam Is Wrong
One of the dumbest sentences intelligent people keep repeating is: “The simplest explanation is usually the correct one.”
The Darija TTS Model Hunt: Failures, False Starts, and a Working Voice
Three text-to-speech experiments, 20.73 hours of synthetic speech, and a working OmniVoice model: lessons in model selection, evaluation, and training on rented GPUs.
What I Would Teach Differently About Machine Learning Today
I reopened an ML/DL course I designed and taught at EMSI. Most of the mathematics survived. My idea of where the hard engineering lives changed considerably.
Getting More Work Out of One H100
Lessons from reducing a multi-model production AI batch pipeline's runtime by 75% on a single H100, covering profiling, I/O, batching, CPU/GPU overlap, model scheduling, caching, backpressure, precision and resilient batch execution.
From Speaker Matching to System Design: Building a Production-Scale Voice Pipeline
What started as a small experiment matching voices across a handful of calls became a local pipeline that backfills tens of thousands of recordings and runs daily, where the hard part stopped being the model and became the system.
How We Expanded DODa to 500K Darija Examples
Darija is inconvenient for NLP: two scripts, flexible spelling, heavy borrowing. DODa’s latest expansion turns that inconvenience into a controlled 500K-row training corpus.
Approach
How I actually think
Measure the uncertainty
A lot of what people call knowledge is borrowed confidence. The useful move is to notice where the chain of justification actually stops.
Read the essays →
Ship conversational systems
Chat, voice, agents. The interesting work is not the demo; it is the system that holds up when a real person starts talking.
See the projects →
Open-source Darija
Moroccan dialect is spoken by millions and still under-served in NLP. Datasets and tools should exist in the open.
Darija Open Dataset →
Write in public
Short notes, long essays, shipped projects. Writing is how the work gets inspected: by you first, then by anyone else.
Browse the notes →
8+
Years in AI
500k
DODa entries
16
Two-minute notes
11
Papers on Scholar