About my Education
Hi, I'm Asim
I'm an AI Engineer based in Cape Town, South Africa. I specialize in Multi-Agent Reinforcement Learning, LLM Agents & Engineering, and Goal-Conditioned Reinforcement Learning
About my Education
I'm an AI Engineer based in Cape Town, South Africa. I specialize in Multi-Agent Reinforcement Learning, LLM Agents & Engineering, and Goal-Conditioned Reinforcement Learning
I'm an AI Research Engineer with experience spanning Multi-Agent Reinforcement Learning and LLM Agents & Engineering. At InstaDeep, I develop MARL algorithms using contrastive learning and curriculum strategies. During my Master's, I built autonomous LLM agents for ML engineering, designed inference-time scaling strategies, and fine-tuned large language models — bridging the gap between RL and modern LLM systems. My recent work, CPPO — the first on-policy contrastive RL algorithm — is under review at NeurIPS 2026 (arXiv:2605.13554).
Currently at InstaDeep working with the Mava team on cutting-edge MARL research. I hold a Master's in AI for Science from AIMS South Africa / UCT — graduated with Distinction (top of cohort) as a Google DeepMind Scholar.
Multi-Agent RL • Goal-Conditioned RL • Contrastive Learning • Curriculum Learning • LLM Agents • Inference-Time Scaling
Python • JAX • PyTorch • vLLM • HuggingFace • TRL • Unsloth • LangGraph • LangSmith • LiteLLM • Hydra • TPU/GPU
Autonomous LLM Agents • Agentic Workflows • Inference-Time Scaling • SFT/Fine-tuning • vLLM Serving • MLE-Bench
Working with the Mava team on multi-agent reinforcement learning research. Developing novel approaches combining contrastive learning, goal-conditioned RL, and curriculum learning for complex multi-agent environments. Also built Tinkerer — an autonomous LLM-powered multi-agent system for automated scientific discovery, featuring a closed-loop workflow of idea generation, code implementation, experiment scheduling, and result collection.
University of Cape Town & AIMS South Africa
AI for Science program in partnership with DeepMind. Focused on reinforcement learning, deep learning, and their applications to scientific problems.
Worked on the development of agriAI, a field monitoring application providing real-time insights into crop health, soil conditions, and environmental factors.
Built a crop classification pipeline using Random Forest with Sentinel-2 multispectral bands, achieving 89.22% accuracy in categorizing key crops in El Gezira. Contributed to the agriAI project — an intelligent farming assistant providing real-time crop health and soil analysis.
A selection of my original research and engineering projects
The first on-policy contrastive RL algorithm, and the first to handle discrete action spaces — learns goal-conditioned behaviour with no reward function, replay buffer, or GAE. Across 18 tasks in 5 suites, matches or beats reward-engineered PPO. arXiv:2605.13554.
Fine-Grained Credit Assignment for RL Training. Token-level reward assignment to improve training stability and sample efficiency for LLMs.
Combining contrastive reinforcement learning with unsupervised environment design for multi-agent curriculum learning.
Autonomous LLM agent for end-to-end ML engineering via tree search. Uses inference-time scaling (Self-Reflection, Planner-Coder, Self-Consistency) and a self-hosted vLLM backend to get a distilled 32B open model to a 30% medal rate on MLE-Bench — matching OpenAI o4-mini. (MSc thesis.)
LLM-powered multi-agent workflow for automated scientific discovery. A closed-loop system where an AI Scientist generates research ideas, an Engineer implements them in code, a Scheduler runs experiments, and a Collector gathers results — autonomously iterating on ML research. Built with LiteLLM, Claude Code, Hydra, and Neptune.
Detecting GenAI-generated content and sophisticated manipulation in public media using machine learning.
Supervised fine-tuning pipeline for DeepSeek-7B. Custom training data curation, LoRA/full fine-tuning experiments, and evaluation for specialized ML engineering tasks.
Neural machine translation system for Arabic to Swahili, addressing low-resource language pair challenges.
Asim Osman, et al. — arXiv:2605.13554, 2026. Workshop version submitted to RLBRew @ RLC 2026.
Asim Osman, et al. — ICLR 2023, Tiny Papers Initiative.
I'm always interested in discussing research collaborations, new opportunities, or just chatting about RL and AI.