Mohamed Rissal Hedna - AI Engineer

Core Stack Python, PyTorch, C/Python Interop, Docker, CI/CD, Git

LLM & GenAI LlamaIndex, Qdrant, Multi-Agent Systems, RAG, Hallucination Detection

Performance Engineering Inference Optimization, Memory Profiling, Quantization, BF16 Training

Research Areas Self-Reflecting LMs, Uncertainty Quantification, Computer Vision

Education

Expected 2026

Thesis: "SERL: SElf-Reflecting Language Models"
CGPA: 1.69 (German Scale, 1.0=Best)

CGPA: 3.91/4.0 (High Distinction) • MEPI TL Full Scholarship Recipient

Jan 2024 – Present

Engineered a high-throughput data pipeline, performing memory profiling to reduce end-to-end latency by 20%.
Designed automated evaluation workflows for multi-agent reasoning systems, including reliability checks and failure case analysis.
Delivered production-grade features driving a 25% increase in user adoption.

Oct 2025 – Jan 2026

Architected an end-to-end generative content engine transforming raw financial datasets into short-form video content.
Led the engineering of a multimodal system (Text-to-Video/Audio), automating the full production lifecycle.
Designed retrieval architecture for complex financial queries ensuring data consistency.

Jan 2023 – Jan 2024

Scaled a multimodal AI platform (PlazmaPunk) to 100,000+ users.
Reduced pipeline latency by 78% by researching and testing efficient vision-language architectures.
Implemented parallel execution strategies to resolve critical high-load bottlenecks.

2022

Engineered a custom C/Python CNN architecture achieving <50ms inference latency on embedded NVIDIA Jetson hardware.
Optimized real-time obstacle detection in GPS-denied environments.

Engineered a custom "T5-Efficient" Transformer for resource-constrained devices.

Slashed parameter count by 81% (17.7M to 3.3M) by optimizing layer distribution.
Orchestrated training in PyTorch using BF16 mixed-precision, converging in under 12 hours on single-GPU.

PyTorch Model Compression BF16

Architected a retrieval-augmented generation system using LlamaIndex and Qdrant.
Optimized retrieval chunks to achieve a 13% reduction in query resolution time.

RAG LlamaIndex Qdrant

SERL: SElf-Reflecting Language Models

M. Rissal Hedna, Chris Biemann. Manuscript in Preparation (Targeting ACL/EMNLP 2025).

Simulated Adversarial Patch Attacks on Vision-Based Logistics Systems

Sesugh Nder, M. Rissal Hedna. Technical Report, University of Hamburg, 2024. arXiv:2511.19254.

A Model for the Automatic Mixing of Multiple Video and Audio Clips

M. Rissal Hedna, et al. Proceedings of the 2023 International Conference on Cyberworlds (CW).