Billal Boumedine

Summary

Embedded ML engineer with 5+ years of experience deploying and optimizing neural networks on resource-constrained microcontrollers. Contributed production code to Espressif's official ESP-DL SDK, including the first YOLO26 deployment on ESP32-P4 and four custom PIE SIMD kernels achieving a 2.54× inference speedup. Strong background in low-level optimization, quantization, and hardware-software co-design.

Experience

AI / Deep Learning Consultant, SARL TECHMOLOGY (MoBot) Jan 2021 – Present

Educational robotics startup · Algeria

Computer vision integration for ESP32-S3-based educational robotics kits (5,000+ learners)
YOLO model optimization and edge deployment for automated object detection challenges
Architecting next-gen platform with NPU-based SoC (1.0 TOPS) for on-device inference

Open-Source Contributions — Espressif ESP-DL

PR #286 - First YOLO26n Deployment on ESP32-P4 [MERGED] Jan – Mar 2026

Deployed YOLO26n (2.4M params) on ESP32-P4 with mixed-precision INT8/INT16 quantization
Built a custom bit-exact INT16 LUT emulator for sensitive activation functions
Reviewed by 4 Espressif engineers over 5 weeks, 80 files
Result: 2,072 ms inference, 0.365 mAP, 1.33× faster than YOLOv11n

PR #310 - Hardware Kernel Optimizations [UNDER REVIEW] Apr – Jun 2026

Developed 4 custom PIE SIMD assembly kernels for ESP32-P4:
TransposePIE: butterfly-pattern register transpose (−303 ms)
NN-LUT: nearest-neighbor lookup with hardware rounding (−156 ms)
HardSiluPie8: fused piecewise-linear activation via bit-shift (−131 ms)
TiledConvBlock: L2-cache-aware fused convolution (−379 ms, up to 13.3× per layer)
Achieved 2,072 → 1,088 ms (1.90× speedup) with 0.361 mAP
Bit-exact validation: 0 errors across 451K tensor values (x86 vs RISC-V)

Open-Source Tools

esp32-p4-jit - PIE SIMD JIT Compiler · 53 ★ on GitHub 2025 – Present

Compile and execute PIE assembly on the ESP32-P4 in under 2 seconds for rapid kernel prototyping

esp_ppq_lut - Bit-Exact INT16 Activation Emulator 2026

Python emulator reproducing ESP-DL's INT16 LUT behavior with zero divergence from hardware

Neural Morphing Engine - Model Transformation Framework 2026

Block-by-block graph transformation with knowledge distillation (512 unlabeled images, no labels)
Quality gates (cosine > 0.99) with automatic rollback; tested 3 transformation types

Education

Master's Degree in Computer Vision, Highest Honors 2024 – 2026

University of Science and Technology Houari Boumediene (USTHB), Algiers, Algeria

Thesis: "Deploying and Optimizing YOLO26n on the ESP32-P4 Microcontroller"

Bachelor's Degree in Information Systems and Software Engineering 2021 – 2024

USTHB, Algiers, Algeria

Technical Skills

LanguagesPython, C, C++, RISC-V Assembly (PIE SIMD) ML / DLPyTorch, TensorFlow, ONNX, YOLO, quantization (INT8/INT16/mixed), knowledge distillation, PPQ EmbeddedESP-IDF, ESP-DL, FreeRTOS, RISC-V architecture, PIE SIMD ISA CompilerMLIR (dialect creation, custom passes), LLVM fundamentals ToolsGit, Linux, CMake, LaTeX, Webots

Activities

CELEC - Electronics & Computer Science Club, USTHB (est. 1987) 2021 – 2024

Built object-tracking robots using YOLO + ESP-CAM with wireless inference pipeline
Hands-on embedded systems, Arduino, IoT, and PCB design

AI Challenges Lead, Algerian Robotics Cup 2022

Languages

Arabic (Native) · English (Technical proficiency) · French (Technical proficiency)

References

Prof. Meriem Sebai

Thesis Supervisor · Informatics Faculty, USTHB

meriem.sebai.ms@gmail.com

Prof. Tarik Zakaria Benmerar

Thesis Co-Supervisor · Electronics Faculty, USTHB

tarik.benmerar@usthb.edu.dz

Prof. Boualem Laichi

Thesis Examiner · USTHB

blaichi@yahoo.com