Kassoum Sanogo Artificial Intelligence, Machine Learning & Data Engineer

Dynamic orchestration of specialized AI agents via a Super Agent (PULSAR) with shared memory and intelligent prioritization.

• Intelligent coordination pipeline
• Shared memory between agents
• Dynamic task prioritization
• Contextual human interventions

In-depth study on algorithmic bias minimization and privacy protection according to European and Canadian standards.

• Explicit consent mechanism implementation
• Data anonymization and pseudonymization
• Model auditability and decision traceability
• International regulation compliance

Data flow traceability, metadata management, and access policy integration based on user profiles.

• Traceable and auditable data pipelines
• Fine metadata and lineage management
• Role-based access control (RBAC)
• Retention and archiving policies

Hisorical document process by creating AI Agent for retranscription

• Analysis and process document pipelines
• Find metadata for context specialisation
• Create AI Platform for this process
• LLM + OCR for relevant process

Applied research on artificial intelligence and plants with API development and data visualization.

• Python and Java API for data collection
• Interactive graphics for visualization
• Multi-platform connectivity
• Botanical data input interface

Implementation of an advanced vectorization system to efficiently process data from complex documents.

• Embedding optimization for different document types
• Improved semantic search accuracy
• Adaptation to multimodal formats (text, images, tables)

Research on optimizing Sentence Transformer embedding models to improve semantic similarity performance.

• Adaptive fine-tuning by domain
• Dimensionality reduction without quality loss
• Model distillation techniques

Development of a persistent memory system for LLMs using SQLite, reducing context loss.

• Intelligent storage of previous conversations
• Dynamic contextual retrieval
• Multi-model adaptation (GPT, Claude, Llama)
• Long-task focus management

Implementation of a comprehensive benchmark system to compare different LLM model performances.

• Standardized performance metrics
• Testing across different domains and languages
• Consistency and creativity evaluation
• Results visualization interface

Semi-automated system for detecting bias or unethical behavior in LLM-generated responses.

• Automatic cultural and social bias detection
• Toxicity and discrimination analysis
• Fairness and equity metrics
• Alert and correction system

Development of local and intrinsic explanation methods for classification model, RAG, or LLM outputs.

• LIME and SHAP implementation for NLP
• Attention visualization techniques
• RAG system explainability
• User interface for interpretation

Defense and filtering strategies against adversarial attacks and prompt injection in AI systems.

• Jailbreak and prompt injection detection
• Intelligent filtering of malicious inputs
• Robustness against adversarial attacks
• Adaptive security mechanisms

Modular design of AI agents with specific capabilities, collaborating via shared protocol.

• Domain-specialized RAG agents
• Web scraping and analysis agents
• Classification and annotation agents
• Modular and extensible architecture

Standardization of agent exchanges via JSON for adaptive message passing protocols.

• Standardized communication protocols
• Adaptive and asynchronous message passing
• Automatic protocol negotiation
• Failure management and recovery

Integration of advanced decision mechanisms for complex task execution with intelligent planning.

• Tree-of-Thoughts and ReAct implementation
• Hierarchical task planning
• AutoGen for automatic coordination
• Global performance optimization

Benchmark on efficiency, consistency, and scalability of collaborative multi-agent systems.

• Collaborative performance metrics
• Role distribution evaluation
• Latency and throughput measurement
• Final result quality and consistency