Mariame Dia

Speech and language AI researcher building ASR, TTS and speech translation for low-resource African languages the exact problem space Intron works in. I have taken a Wolof ASR system from an empty dataset to production: 150-hour corpus construction, Whisper fine-tuning with LoRA/PEFT, WER-driven evaluation, then a CTranslate2 float16 inference migration validated against the incumbent model before any traffic switch. I also run a Wolof StyleTTS2 text-to-speech API in production.

Doctoral researcher (UCAD/ESP, LIMBI) on LLM systems for Wolof, Pulaar and Hassaniya.

Native

Pulaar speaker, fluent in Wolof, Arabic, French and English I build for these languages and I speak them.

Data Scientist & AI Engineer at CONCREE (2026-01 – Present)

Reference: abdoul.sy@concree.com

• Owned model evaluation protocols defined the metrics, robustness and safety tests and acceptance thresholds used to decide whether a model ships, and arbitrated the resulting go/no-go calls.
• Migrated the production Wolof ASR from Transformers fp32 to faster-whisper (CTranslate2) fp16, aligning the decoder configuration (greedy beam_size=1, no forced language token, temperature 0, thresholds disabled) so that outputs stayed comparable. Transcriptions were identical to the incumbent model on the validation set; container size dropped 40% (10.8 GB → 6.5 GB) and cold start fell accordingly. Built a WER guard-rail script that blocks the switch above a 1% relative WER drift.
• Deployed and optimised a Wolof TTS API (StyleTTS2, 5 voices, voice cloning, streaming endpoint) on GPU serverless infrastructure; diagnosed the latency budget and cut model load time from ~63 s to 11.4 s via a persistent network-volume weight cache, with first audio byte at ~0.8 s on the streaming route.
• Generative AI and applied NLP: fine-tuned Qwen on domain data, built RAG pipelines for synthetic data generation, benchmarked frontier LLMs for production selection, and shipped a WhatsApp conversational agent that lets non-French-speaking users enrol in and follow courses on the Lezgo platform.

AI & Speech Engineer (internship → full role) at CONCREE (2025-07 – 2025-12)

• Built an end-to-end ASR pipeline for Wolof, a low-resource language with no usable public corpus: dataset construction, preprocessing, Whisper fine-tuning, WER evaluation and inference optimisation.
• Assembled the speech corpus from scratch (150 hours) by automating collection with yt-dlp and the YouTube Data API, then extracting subtitles, aligning audio to text, denoising and running quality control deliberately sampling across speakers and registers to cover accent variation.
• Applied parameter-efficient fine-tuning (LoRA / PEFT) to make repeated fine-tuning runs affordable on constrained GPU budgets, and documented the pipeline for reproducibility and hand-over.

Speech Translation Engineer TEKKIA project at École Supérieure Polytechnique (ESP) (2024-01 – 2024-04)

Reference: demba.faye@esp.sn

• Built a Wolof→French speech translation pipeline: web collection, audio segmentation, noise removal, transcript alignment and corpus construction; prepared 20 hours of speech as ~13,056 aligned segments.
• Fine-tuned and evaluated models using WER, BLEU and chrF, and delivered an automatic subtitling application driven by Wolof speech input.

Co-Founder & Technical Lead at Learniverse (2023-01 – Present)

EdTech platform

• Lead development of an offline-first educational mobile app with an LLM tutoring assistant that reads PDF and image course material and explains it interactively.

MSc in Data Science (Mathematics) – African Institute for Mathematical Sciences (AIMS) (2024-01 – 2026-01)

MSc in Artificial Intelligence & Big Data – École Supérieure Polytechnique (ESP) (2023-01 – 2025-01)

BSc in Telematics Software Engineering & Network Administration – ESTM (2019-01 – 2022-01)