Security Agent
Fine-tuned Qwen3.6-27B model trained for cybersecurity tasks and bug finding.
RESEARCHER · BUILDER
Ibra Niang
Multimodal NLP researcher at the University of Memphis (BS/MS '26), building vision-language models and the systems that train them.
Selected Work
Things I've built or am building. Code and write-ups linked.
Jaxchat
End-to-end LLM training pipeline: pretraining, SFT, RL, and a chat web UI for inference, all in JAX.
Multimodal Training
CUDA/C multimodal training stack for multimodal LLMs across image and vision modalities.
Agriculture VLM Benchmark
Research-grade benchmark for vision-language models in agriculture: dataset, evaluation harness, and reference baselines.
in progress · paper forthcoming
Writing
Notes on VLMs, training, kernels, and building research tooling.
From ~0.15 to ~0.60 reward: fast RL gains on low-resource translation
How curriculum, reward shaping, and decoding stability moved reward quickly in a low-resource translation setup.
all writing →Get in touch
Open to research collaborations, applied roles, and conversations.