Siham Janati
Siham Janati

Siham Janati

AI software engineer · full stack & mobile

I build AI products people trust with real work: clinical notes written from a dental visit, schedules that respect every constraint of a busy practice, and an app that brings a classical book to life in sound. From the first sketch to production.

Master's in Data Science, StrasbourgBased in MoroccoArabic · French · English

See the work Get in touch
Speech · text · time
الشفا00:00.02
بتعريف00:00.30
حقوق00:01.14
المصطفى00:01.72
ﷺ00:02.46
A thread through my work: dental visits transcribed live, recitations aligned word by word.

Recited by El Mostafa Belhiad, from the al-Shifa app.

Selected work

Products in production

2024 — now End to end: product, design, engineering and AI With Hamza El M'houjab and Hicham Janati kailavoice.com ↗

Kaila Voice

AI clinical notes for dental practices. The team records the visit; Kaila transcribes it live and writes the note, which staff review and post straight into Open Dental, the practice system they already use.

  • Built end to end, from the product and its screens to the web app, the backend, the AI and the weekly releases to dental offices in production.
  • The notes: SOAP notes, templated notes, visit summaries and patient briefs, with rules that keep tooth numbers consistent across sections, repair speech-recognition slips in dental vocabulary without ever touching a number, and state each fact once.
  • Each office chooses its AI model with a cost estimate, and every note records which model wrote it.
  • We treat patient data with care: personal and health information detected, logs stripped of it, retention limits, agreements accepted in the app.
ReactPython · FastAPIAWSOpenAIPresidio · spaCyStripe
2025 — now End to end: product, design and engineering Live beta with pilot practices With Hamza El M'houjab dentalai.io ↗

Dental AI

Scheduling and online booking for dental practices, built on top of Open Dental: it finds the right time for every procedure and lets patients book online, while Open Dental stays the system of record.

  • A scheduling engine that times each procedure as setup, provider and assistant time, searches slots across providers, rooms, blockouts and daylight-saving changes, and ranks what it finds.
  • Two-way sync with Open Dental: live change events in, hourly reconciliation, and a careful, reversible switch for offices that make it their schedule source.
  • A patient hub for the front desk (upcoming visits, recalls, pending treatment, the whole household on one screen) and a booking site for patients.
  • Built end to end: the scheduling engine and backend, the staff app, the admin panel and the patient booking site.
Java 17 · Spring BootPostgreSQL · FlywayReact · TypeScriptAWS
2025 — now Product, design and engineering, end to end Not-for-profit alshifaa.app ↗

Al-Shifaa الشفا

Qāḍī ʿIyāḍ's twelfth-century classic on the Prophet ﷺ, as an app to read and to hear: free, no ads, no accounts.

  • A pipeline that turns a printed edition into structured, verified text, then places a full recitation on it with speech recognition and forced alignment, so every line lights up as it is read.
  • A reader, a read-along with the voice, and an immersive mode that follows the recitation clause by clause.
  • French and English translations with synthetic narration, every line reviewed by a person before it reaches readers.
  • An admin platform behind it: reader reports and replies, in-app notices with A/B tests, anonymous usage statistics, translation review.
69recorded episodes, aligned
168sections of verified text
3interface languages
2.0on the App Store and Google Play
Flutter · DartPythonTypeScriptCloudflare Workers · D1 · R2WhisperElevenLabs
2026 Speech alignment Timing data for a devotional app

Word timing for chanted texts

Line-by-line and word-by-word timing for hours of melodic recitation (Dalāʾil al-Khayrāt, al-Hamziyya, al-Burda), where ordinary speech tools lose their way.

  • Replaced a Whisper pipeline with CTC forced alignment on a model trained on Qur'an recitation, cross-checked against Meta's MMS, with review pages for checking by ear.
  • In one recording, 458 of 483 lines and 193 of the 201 names of the Prophet ﷺ placed automatically; all 174 lines of the Burda.
  • Traced timings that drifted in players to variable-bitrate MP3 seeking, and fixed it at the source.
PyTorch · torchaudiowav2vec2 CTCMeta MMSffmpeg
2026 Concept, direction and edit Short film · 1:47 Watch on Instagram ↗

After the Game من بعد الماتش

A short film for Al-Shifaa's 2.0 launch: friends from different countries, a football match, a dream, and a book that finally speaks their language.

  • Made with generative video, then voiced with cast voices timed to each mouth, keeping the crowd and room sound of the scenes.
  • Bilingual captions and message cards designed for the story, an intro and outro in the app's identity, mixed to broadcast loudness.
Generative videoVoice designSource separationffmpeg
Earlier work

Machine learning, research and systems

Fraud detection for card payments

Hightech Payment Systems, Casablanca · 2023

Machine-learning models to improve a rule-based payment-fraud system: data generation, the full analysis pipeline, models trained in Python and served in Java through PMML and ONNX.

Segmenting satellite images

ICube laboratory, Strasbourg · 2022

A new segmentation algorithm for remote-sensing images based on the Normalized Cut criterion, in Java, and a DBSCAN variant for spatial clustering.

A compiler from scratch

Université de Strasbourg

A compiler for the Decaf language in C with Flex and Bison, emitting MIPS assembly.

A chatbot for print requests

Université de Strasbourg

BERT to sort valid requests from the rest, spaCy to read their details, a scheduler and a dashboard.

Malware and anomaly detection

Université de Strasbourg

Obfuscated malware caught at runtime with SVM, Random Forest and XGBoost; time-series anomalies on oil-well data with a 1D convolutional network.

Mining GenBank, and two games

Université de Strasbourg

Thousands of genome files mined in parallel behind a PyQt interface; Atari's Tempest rebuilt in modern C++ and SDL; the card game BANG! as a multiplayer Unity game.

Capabilities

From the first sketch to production

AI products

LLM features that hold up in real work: prompts and pipelines with domain rules, model choice and cost, outputs people review.

Speech & audio

Live transcription, forced alignment, synthetic voices, source separation and mastering.

Full-stack web

React and TypeScript front ends; Python, Java and TypeScript backends; integrations with systems of record.

Mobile

Flutter apps for iOS, Android and the web, released and maintained in both stores.

Cloud & data

AWS and Cloudflare, PostgreSQL and DynamoDB, careful migrations, health-data hygiene.

Data science

Classical ML and deep learning, from feature work to models served in production.

Experience & education

The road so far

  1. 2022 — nowFreelance AI software engineerFull stack and mobile, for clients and for my own products
  2. 2023Machine learning internHightech Payment Systems (HPS), Casablanca
  3. 2022Research internICube laboratory, Strasbourg
  4. 2016Graphic design internDyar al Mansour, Rabat
  1. 2021 — 2023Master's in Computer Science, Data Science & Complex SystemsUniversité de Strasbourg · with honours
  2. 2019 — 2021Bachelor's in Computer ScienceUniversité de Strasbourg
  3. 2017 — 2019Mathematics & Computer ScienceUniversité Mohammed V, Rabat
  4. 2015 — 2016Graphic & interior designArt'com Sup, Rabat · top of the class
Approach

How I work

Fidelity first

A clinical note, a schedule or a sacred text is a trust. Every output is checked against its source before anything clever is built on it.

People decide

Models draft, align and suggest. A person reviews, chooses and signs off.

Calm by default

No noise, no dark patterns: products people return to because they help.

Beyond the screen

Colour, books and the road

Painting & drawing

The designer came before the engineer, and still shapes every screen.

Reading

Books in Arabic, French and English, the old and the new.

Travelling

Four years of study in France, and the habit of seeing a place, and a problem, from somewhere new.

Contact

Have a product that needs to think, listen or read?

AI features people can trust, speech and audio pipelines, Arabic-first products, mobile and web apps. Tell me about yours.

contact@sihamjanati.com LinkedIn ↗ GitHub ↗