Minh Lê Đức

QUAVEO

To be updated.

Apollo Solutions VN

• Internal tooling: Built a multimodal agent for internal tooling.
• Social listening: Built a multimodal agent for social listening.
• Profiling at scale: Social media profiling at scale.

Oryza Systems

• AI services: Designed and deployed real-time face recognition systems for secure, low-latency identification across edge and cloud environments.
• Vector retrieval: Integrated Faiss and Milvus to enable high-speed vector similarity search and information retrieval for image recognition.
• Deployment: Built containerized inference services using Docker and led the entire deployment lifecycle, including CI/CD, monitoring, performance tuning, and rollback strategies.
• Metadata services: Designed and implemented scalable metadata pipelines for ingestion, enrichment, and system-wide synchronization using Kafka.
• Infrastructure: Managed microservices with Docker and Kubernetes, and improved system resilience with logging and monitoring via the ELK stack.

HPT Vietnam Corporation

• Model development: Created an AI-powered malware detection system with 95% accuracy, reducing false positives by 30%.
• Optimization: Enhanced model efficiency and accuracy, reducing inference time by 25%.
• Documentation: Produced detailed reports for knowledge transfer and model reproducibility.

Kalman filtering and state estimation — ten filter families under one consistent API.

• Breadth: Kalman, Extended, Unscented, Particle, Ensemble, Information, Adaptive, and Square-Root filters, plus an RTS smoother.
• Tracking: Multi-object tracking with Hungarian association and ready-made motion models.
• Numerical rigor: Joseph-form updates and Cholesky stabilization, with RMSE, NEES, and NIS diagnostics.

import numpy as np
from kalbee import KalmanFilter

kf = KalmanFilter(
    np.zeros((2, 1)),             # state
    np.eye(2),                    # covariance
    np.array([[1, 1], [0, 1]]),   # F: constant velocity
    np.eye(2) * 0.01,             # Q
    np.array([[1, 0]]),           # H: observe position
    np.array([[0.1]]),            # R
)
kf.predict()
kf.update(np.array([[1.2]]))
# kf.x -> [[1.14], [0.57]]

PyPI Source

Rust-backed swarm intelligence optimization for Python — a Pythonic interface over a compiled core.

• Eleven algorithms: Ant Colony, Particle Swarm, Artificial Bee Colony, Grey Wolf, Firefly, Simulated Annealing, Cuckoo Search, Bat, Glowworm, Bacterial Foraging, and Differential Evolution.
• One entry point: A unified AutoColony API covers continuous optimization and discrete problems such as the Traveling Salesman.
• Rust core: Compiled via PyO3 and shipped to both registries.

from colonyx import AutoColony

def sphere(x):
    return sum(xi * xi for xi in x)

opt = AutoColony(
    mode="pso", n_iterations=150, random_state=42
)
opt.fit(sphere, bounds=[(-5, 5)] * 3)

opt.predict()   # -> ~[0, 0, 0]
opt.score()     # -> ~0.004

PyPI crates.io Docs

A high-performance Enigma machine simulator covering the M3 and M4 models, with Python bindings for anyone who would rather explore the cipher from a notebook.

• Historical fidelity: Models the rotor, reflector, and plugboard mechanics of the M3 and M4 machines.
• Rust core, Python bindings: A Rust implementation exposed to Python through PyO3, for teaching, tinkering, and cryptanalysis experiments.

use enigmar::EnigmaBuilder;

let mut machine = EnigmaBuilder::new()
    .rotor("I", 0, 0)      // type, position, ring setting
    .rotor("II", 0, 0)
    .rotor("III", 0, 0)
    .reflector("B")
    .plugboard("AV BS CG DL FU HZ IN KM OW RX")
    .build()
    .unwrap();

let ciphertext = machine.process_string("HELLOWORLD");

crates.io Source

An AI-powered brand analysis platform that processes multimodal data — text, images, audio, and video — to evaluate brand perception and sentiment.

• AI-driven insights: Led a team of 4 to develop a brand-awareness evaluation system leveraging multimodal deep learning.
• Cloud scalability: Integrated Google Cloud AI services for real-time processing and high availability.
• Impact: Achieved 90% accuracy in brand mention detection and 85% sentiment classification accuracy.

• Streamlit
• Python
• Google Cloud
• Ngrok

A deep learning project exploring AI-assisted music composition and ideation using neural networks.

• AI composition: Created a system that generates music using PyTorch-trained neural models.
• Scalable APIs: Deployed via FastAPI for real-time inference and integration.
• Creative impact: Contributed to research on AI-enhanced creative workflows in music production.

• FastAPI
• PyTorch
• MongoDB

Writing about AI systems, generative models, MLOps pipelines, and real-time vision — in both English and Vietnamese, to make technical material reach further than it usually does.

• English
• Vietnamese

Bachelor’s Degree, Faculty of Computer Science

Binh Duong Province, Vietnam

Bachelor’s Degree, Faculty of Computer Science

Frankfurt, Germany