Software & machine-learning engineerCTO, MYTEAM.aiParis region

Mouadh Bouali

I build the systems behind applied AI, from the model to the operator’s screen.

  • Product engineering
  • ML infrastructure
  • Applied AI
  • Research

Software, end to end.

I’m an engineer near Paris and CTO of MYTEAM.ai, a small AI company where I still write and review the code. I lead a video-analytics platform from its event store to its Kubernetes and Jetson deployment paths, build LLM products in which a person stays in control, and run the research that decides what is ready.

See selected work
Role
CTO / Responsable technique, MYTEAM.ai, since June 2026
Works across
Product engineering, ML infrastructure, applied AI, research
Builds in
Python, TypeScript, C#
Platforms
PostgreSQL, Docker, Kubernetes, ArgoCD, NVIDIA Jetson
Trained at
Télécom SudParis, Institut Polytechnique de Paris: engineering degree, 2025
Languages
Arabic, French, English · German (B1)
Based in
Cachan, Paris region, France
Selected work · 05 entries

Selected work

Three products and two research programs, each with my role and where it stands today.

01Video-analytics platformMYTEAM.ai

Avertix

Camera streams in, reviewed alerts out.

Event and alert processing on FastAPI and PostgreSQL, a React console for operators, and one inference worker packaged for cloud GPU nodes and for Jetson devices beside the camera, delivered through CI gates, ArgoCD and canary releases.

Role
Hands-on technical lead
Status
Company platform; GPU demonstration environment operated; Jetson path packaged
Built with
Python, FastAPI, PostgreSQL, React, TypeScript, Docker, Kubernetes
Runtime path
  • Led by me
  • Outside the platform
  1. Camera sourcesAdded and diagnosed from the console
  2. Inference workerOne codebase, packaged for cloud GPUs and for Jetson on site
  3. Control planeFastAPI on PostgreSQL; immutable, versioned events
  4. Alert projectionsWritten in the same transaction; acknowledge lifecycle
  5. Avertix ControlReact and TypeScript console for operators

02Operator cockpitLocal-first application

Ana Concierge OS

The assistant proposes. A person approves.

A local-first app for case work, approvals and governed knowledge. Each assistant turn is rebuilt from the database, may cite only passages it actually retrieved, and turns any action into a proposal that waits for a person.

Role
Sole developer
Status
Working local foundation; no outbound actions by design
Built with
TypeScript, React, Fastify, PostgreSQL, zod, Vitest
One assistant turn
  • Built by me
  • A person or a model
  1. An operator asks about a casePerson
  2. Context rebuilt, passages retrieved in scopeAna server
  3. A model drafts schema-valid JSONBounded runner: no shell, no general tools
  4. Citations and scope checked; proposal hashed and queuedAna server
  5. A person approves or rejects; the decision is auditedPerson

03Group-booking assistantMYTEAM.ai

Livia

Livia answers restaurants’ group-booking requests on WhatsApp. I work on its conversational agent as part of the MYTEAM.ai team.

Role
Engineer on the conversational agent, in a team
Status
Live product

Research

Two programs, each with a public overview

04Collaborative researchTransFlower and V-JEPA

Video models

Two branches, one decision, then a test away from home.

TransFlower, a dual-branch Transformer-CNN model accepted for oral presentation at KES 2026, pairs a VideoMAE Transformer branch with an RGB and optical-flow I3D branch. I was responsible for much of the Transformer branch, developed the I3D branch and integrated the two into one inference path. I then ran the later experiments with frozen V-JEPA representations on footage from cameras outside the training data.

Role
Co-author: Transformer branch, I3D branch, hybrid integration
Status
Paper accepted at KES 2026; external results exploratory
TransFlower: two branches, one decision
  • My work
  • Pretrained

Transformer branch

  • Pretrained VideoMAE encoder
  • Token pooling with learned attention
  • Attention across clips
  • Adaptive classification head

Motion branch

  • RGB and optical-flow I3D
  • Hybrid inference pathBoth branches combined into one decision

Later campaign

  • Frozen V-JEPA representations
  • Readouts tested on footage from other cameras

05Independent researchTabular self-supervised learning

T-JEPA

Make the collapse measurable before changing it.

Collapse diagnostics and residual reweighting built on an open T-JEPA implementation. The change moved the escape from collapse earlier in all six paired comparisons; downstream quality stayed mixed, so the study reports a timing result, not a quality gain.

Role
Independent researcher, extending the upstream method
Status
Research; public overview online
The study
  • My work
  • Upstream
  1. Open T-JEPA implementation
  2. Collapse diagnostics on dynamics and geometry
  3. Residual reweighting of the objective
  4. Paired comparisons, with controls
Capabilities · five layers

Inside the stack

Grouped by what each layer makes possible, with the work where it was done.

Five layers, from what a person sees down to how it ships. Each has its own pattern; the layer you are reading lights up.
01

Interfaces

People can see what a system is doing, and act on it.

  • Operator consoles in React and TypeScript, covered by component, end-to-end and automated accessibility tests. Avertix
  • Assistant interfaces that show citations, uncertainty and pending approvals instead of hiding them, with a synthetic mode for exploring safely. Ana

ToolsReact, TypeScript, TanStack Query, zod, Vitest, Playwright, axe; Angular (2023)

02

Systems

Behavior stays correct when one of the components is a language model.

  • Event-sourced storage: immutable, versioned events, with alert projections written in the same transaction. Avertix
  • Alert delivery to external applications through signed webhooks, a durable outbox with retries and replay, and adapters that must pass conformance checks. Avertix
  • Server-side boundaries for assistants: context rebuilt on every turn, unretrieved citations and out-of-scope actions refused. Ana
  • Typed services in FastAPI and Fastify over PostgreSQL. Avertix, Ana

ToolsPython, FastAPI, TypeScript, Fastify, C#, .NET, PostgreSQL; Spring Boot (2023)

03

Models

Model choices rest on evidence, not on the first attractive curve.

  • Adapting pretrained video Transformers: attention pooling over tokens, attention-based aggregation across clips, adaptive classification heads; an I3D branch and a hybrid. Video models
  • Mechanism diagnostics and objective interventions in self-supervised training. T-JEPA
  • Running models where the data is allowed to be: one inference worker packaged for cloud GPUs and Jetson devices. Avertix
  • Model conversion and quantization experiments on a T4 GPU (ONNX, INT8, TensorRT), compared on file size, accuracy and inference time. Avertix
  • LLM assistants with structured output, bounded tools and human approval. Ana

ToolsVideoMAE, I3D, V-JEPA, T-JEPA; structured LLM output

04

Data & evaluation

A result can be trusted because the data and the comparison can.

  • Content hashing that exposed duplicate videos across a train/holdout boundary, followed by retraining on clean splits. Video models
  • Controls that are allowed to say no: zero-weight and shuffled-label runs, predeclared criteria, candidates frozen before external data. Video models, T-JEPA
  • Retrieval with scope, provenance and checked citations on PostgreSQL full-text search. Ana

ToolsPython, SQL, PostgreSQL full-text search, content hashing, paired comparisons

05

Delivery

Changes reach running systems in small, checked steps.

  • Merge gates: type checks, import-layer contracts, tests, secret and dependency scans, independent review, green checks on the exact commit. Avertix
  • Kubernetes delivery configured with ArgoCD, canary rollouts, a GPU pool that scales to zero, and pull-based updates for Jetson devices. Avertix
  • Containers with GPU and CPU profiles; lazy, thread-safe model loading with an asynchronous warm-up. Avertix

ToolsDocker, Kubernetes, ArgoCD, Argo Rollouts, Rancher Fleet, mypy, import-linter, pytest, gitleaks, OSV

Practice · four habits

How the work gets done

The same habits in products and in research: find the failure, narrow what the model decides, and make every change pass something that could have stopped it.

  1. 01

    Reproduce before repairing.

    A defect becomes a failing test before it becomes a fix. A suspected training failure becomes a diagnostic before it becomes an intervention.

  2. 02

    Give the model a narrow job.

    Models propose; validated state, rules or a person decide. That boundary lives in code and tests, not in a prompt.

  3. 03

    Merge through gates.

    Typed layers, tests, security scans and independent review, with merges allowed only when the checks pass on the exact commit.

  4. 04

    Keep the corrections.

    When a control disagrees with an attractive result, the result goes, and the record says why.

Corrections on record

Results that a control changed or bounded.

Revision A
T-JEPA
An explanation dropped. A register token was not required for escape from collapse in the tested setting, so the study moved to measuring the transition itself.
Revision B
T-JEPA
A timing result, not a quality gain. Residual reweighting moved the transition earlier in all six paired comparisons; downstream accuracy stayed mixed.
Revision C
T-JEPA
A result withdrawn. A representation-subspace finding did not survive a corrected setup and was retracted.
Revision D
Video models
Leakage removed. Hashing every video exposed exact duplicates crossing the train/holdout boundary; the hybrid model was retrained on clean splits.
Revision E
Video models
A candidate rejected. Readouts were frozen before any external footage was used, and a site-adaptation method that missed its predeclared criteria was dropped.
Record

Experience and education

MYTEAM.ai is a small company, so technical leadership there is hands-on.

As CTO I make the architecture calls, set the delivery protocol (issue-scoped branches, independent review, merges gated on passing checks) and review the code that ships. I was lead and sole reviewer for two interns’ feature work, and I remain the hands-on technical lead of the Avertix platform.

  1. Jun 2026 – now

    MYTEAM.ai · Paris region

    CTO / Responsable technique

    Technical lead of a small team: architecture, delivery protocol, code review and mentoring. Lead of the Avertix platform; contributor to Livia.

  2. Dec 2025

    MYTEAM.ai

    Research Engineer, permanent position

    Product and research engineering.

  3. May – Oct 2025

    MYTEAM.ai

    Research internship

    Machine-learning research.

  4. Jul – Aug 2023

    VERMEG · Tunis

    Software engineering internship

    Back-end and front-end work with Spring Boot and Angular.

Education

  • 2021 – 2025

    Télécom SudParis, Institut Polytechnique de Paris

    Engineering degree (diplôme d’ingénieur), Statistical Modeling and Machine Learning

  • 2018 – 2021

    Lycée Janson-de-Sailly, Paris

    Classes préparatoires, MPSI then MP*

Publication

  • 2026

    TransFlower: Contextual Dual-Branch Transformer-CNN Architecture for Real-Time Violence Detection

    Co-author · accepted for oral presentation at KES 2026. Research overview

Languages

  • Arabic (native), French (C2), English (C2), German (B1)

Outside work

  • Music, piano and video editing

Contact

Tell me what you’re building.

Email is the best way to reach me. I work in English and French.

mouadhbouali@gmail.com