Skip to content
Deiss Yehouenou.

Work

Turning data into decisions you can defend.Measure

Pilotage financier — Société Générale

  • In production
  • At work

A database centralising direct, indirect and variable costs across business units, plus profitability dashboards used by financial controllers.

What I did
  • Several levels of granularity in one report: a controller moves from the overall view to a single cost line without switching tools.
  • Internal process automation following a Lean approach.
  • SQL
  • Power BI
  • DAX
  • Excel
  • Lean

Scoring explicable — PGS Africa

  • MVP
  • At work

An analytics MVP that automatically categorises at-risk cases from event data, and can say why.

What I did
  • Explainable by design: a flagged case comes with the elements that raised it, otherwise nobody uses it.
  • Industrialisation of the related reporting: ETL pipelines, DAX measures, operational indicators.
  • Python
  • Power BI
  • DAX
  • Power Query
  • ETL

Reconstitution tarifaire

  • In production
  • For a client

Rebuilding a catalogue's complete price history from analytics events alone, to verify no customer ever paid an inconsistent price.

What I did
  • 1,715 price versions observed across 1,526 combinations of service, variant and options.
  • 43 combinations whose price actually changed, isolated with their dates and intermediate values.
  • 155 control bookings whose price was recomputed to check it matched the catalogue in force that day.
  • PostHog
  • JSONL
  • Python

Simulation Monte-Carlo d'activité

  • Prototype
  • Personal project

A probabilistic model of a service business's revenue, to reason in ranges rather than a single projection.

What I did
  • A single forecast gives false certainty; a distribution says what can actually happen, and how likely it is.
  • Python
  • Monte-Carlo

Impact des actualités sur les marchés

  • Prototype
  • Personal project

A complete pipeline that collects financial news feeds, links them to market prices and trains a model to estimate their effect.

What I did
  • Syndication feed collection, market prices, relational storage, service API and a reading dashboard.
  • Python
  • FastAPI
  • scikit-learn
  • SQLAlchemy
  • Streamlit
  • yfinance

Complete web products, shipped and used.Build

HairByGloria

  • In production
  • For a client

A complete booking platform for a salon: customer site, booking funnel, management back-office and built-in AI agents.

What I did
  • Admin-managed catalogue where variants and options change both price and duration, availability computed by interval, configurable opening hours and closures.
  • The server recomputes every price before committing it: no value coming from the browser is trusted.
  • Error tracking, consented analytics, images converted on upload, admin authentication without an external dependency.

hairbygloria.fr

  • Next.js
  • Supabase
  • PostgreSQL
  • Sentry
  • PostHog
  • bcrypt
  • sharp
  • Vercel

deiss.website

  • In production
  • Personal project

The site for my digital engineering practice: a service catalogue, email enquiries, and a public tool that audits the tracking of any third-party page.

What I did
  • Strict consent gate: a refusal produces complete silence, including for the built-in demo.
  • A public diagnostic inspects a third-party page's tracking without executing its JavaScript — requests hardened against internal redirection attacks, nothing stored.
  • Prices and labels come from a single catalogue: an expiring offer disappears everywhere by changing one date.

deiss.website

  • Next.js
  • TypeScript
  • PostHog
  • Resend
  • Vercel
  • Meta Pixel

Klassic Horlogery

  • MVP
  • Personal project

A commercial landing page with lead capture and full attribution: UTM, Pixel, GA4, Clarity.

What I did
  • The form records where the visitor came from: without it you know how many leads arrive, never from where.
  • Next.js
  • Supabase
  • Meta Pixel
  • GA4
  • Clarity

Festichill — billetterie de festival

  • In production
  • For a client

Online ticketing for a festival in Cotonou: a three-screen purchase funnel, limited stock, mobile or cash-on-delivery payment, e-ticket by messaging.

  • 300+tickets sold
What I did
  • Guest checkout: three screens from choosing a ticket to confirmation, designed for one thumb on a phone.
  • Two ticket formats, quantity discounts and delivery fees by zone — the total is recomputed server-side, never accepted from the browser.
  • Real stock shown honestly rather than a fake urgency counter: selling the same ticket twice costs more than losing a sale.
  • A back-office built to be run from a phone, because that is how the seller actually works.

festichill.com

  • Next.js
  • Supabase
  • Zod
  • PostHog
  • Playwright
  • Vitest
  • Vercel

Getting systems to do the work, with guardrails.Automate

Agents de relation client

  • In production
  • For a client

Conversational agents wired into Messenger and Instagram: they qualify the request, check genuinely available slots, and hand over to a human when needed.

What I did
  • The agent reads real availability before offering a slot: it does not guess, it queries the system.
  • Explicit escalation rules — a complex case goes to a human rather than getting an approximate answer.
  • Automated customer follow-ups, keeping the conversation context.
  • OpenRouter
  • Meta Graph API
  • Webhooks
  • Next.js

AI Software Factory

  • Prototype
  • Personal project

An agent-driven software production pipeline, framed by explicit contracts and verified against reference tasks.

What I did
  • Each task is bounded by a contract: an agent that strays outside the scope fails verification.
  • A disposable sandbox repository and reference tasks with a known initial state, to measure the loop on a real case rather than a demo.
  • Python
  • Pydantic
  • YAML
  • Node.js

Chaîne de candidature vérifiée

  • In production
  • Personal project

The system I use to find my apprenticeship: a CV tailored to each offer, generated from a single source of facts, which cannot be sent until it clears a blocking check.

What I did
  • A single source of facts, justified line by line: every profile item carries its origin and date. A CV is never written by hand, it is regenerated from that source.
  • A blocking six-part final check: factual accuracy, file fingerprints, spelling, coverage of the advert's keywords, and a comparison of every rewritten sentence against its original — the system flags responsibility that grew in translation, such as "contributed to" becoming "led".
  • The approved PDF is frozen by its fingerprint: any regeneration voids the approval, so the file that goes out is exactly the one that was reviewed.
  • Before tailoring a CV, the system asks targeted questions about what the advert wants and the profile does not yet show. It never infers a skill: a requirement left uncovered does not enter the CV.
  • One folder per application, frozen once sent, keeping the advert, the requirement-to-evidence mapping and the exact version submitted — that is what the interview is prepared from.
  • Python
  • python-docx
  • pdfminer
  • Google Drive
  • Google Sheets
  • Claude Code

Génération vidéo automatisée

  • Prototype
  • Personal project

A pipeline that composes and renders short videos with no manual step: animation, capture, encoding.

What I did
  • Code-driven animation, frame-by-frame capture in a headless browser, automated final encoding.
  • Node.js
  • GSAP
  • Puppeteer
  • FFmpeg

Things I built because the question interested me.Off-piste

Audits de sécurité de mes propres applications

  • In production
  • Personal project

Penetration test reports run against my own web applications, and a reusable audit procedure that came out of it.

What I did
  • Auditing what you wrote yourself is uncomfortable: that is exactly why it is worth doing.
  • Sécurité applicative
  • OWASP

Réalité virtuelle

  • Academic project
  • Studies

Interactive virtual reality environments built during my engineering studies.

What I did
  • Hand interaction, voice recognition, and the constraint that shapes everything else: holding the frame rate, or you make the user sick.
  • Unity
  • C#
  • Meta XR SDK
  • Oculus

Jeu 3D dans le navigateur

  • Prototype
  • Personal project

A three-dimensional game running in a browser, with a physics engine and spatial audio.

What I did
  • Real-time rendering, physics simulation and audio, with nothing to install on the player's side.
  • Three.js
  • Rapier
  • Howler.js

L'Expédition

  • Prototype
  • Personal project

A 21-day anti-procrastination programme: daily missions, rewards, and conversational coaching.

What I did
  • Scheduled tasks, instant-messaging notifications, state kept in a database — built for myself first.
  • Node.js
  • Express
  • PostgreSQL
  • Telegram Bot API
  • Anthropic