Architecture and development accounts. Not a benchmark leaderboard.
Conceptual associations · not experimental data
An interactive reading guide
Explore the connections.
Choose a topic to inspect its described methods. Lines show thematic associations, not citation links, measured outcomes or validated dataflow.
The full research accounts below are available without an interactive map.
University–industry research intelligence
Architecture of an integrated platform linking researchers, students, companies, research activities and industry needs at Amirkabir University of Technology.
What the CV describes, which components are planned, and what the available evidence does not establish.
01
Research data
University–industry research intelligence
Connecting research capabilities with technological needs.
Architecture of an integrated platform linking researchers, students, companies, research activities and industry needs at Amirkabir University of Technology.
Architecture described; intelligent services are a roadmap
Data provenanceEntity resolutionSemantic retrievalKnowledge graphs
Described work
A unified data-management layer: source registry, staging, metadata catalog, lineage, quality checks, entity resolution, master data and synchronization operations.
A shared model for people, organizational units, theses, projects, grants, technologies, companies, laboratories, courses and technology–market relationships.
A central API gateway for controlled data exchange rather than direct point-to-point integration.
Analyst decisions, alert status and written explanations become new training information for later alert classification.
Explainable alert views expose scoring reasons, influential features, feature importance, similar events, review status and model-training information.
Ingestion is described for CSV, event logs, database outputs, multiple data sources and direct database connections, with detection of newly available files.
A natural-language scenario builder translates a requested scenario into rules, conditions and executable alerts; the architecture separates application logic from the language-model provider.
Technical components
DjangoOllamavLLMCompatible model APIs
Planned / roadmap
Continuous evaluation using accuracy, precision, recall, F1, false-positive and false-negative rates, and change in performance over time.
The metrics are evaluation provisions, not achieved scores. The CV does not provide a dataset, benchmark result or quantified reduction in fraud.
Source: supplied AI portfolio · pp. 3, 4, 5
03
Agent systems
Tool-connected analytical agents
Linking a request to tools, data and a structured analytical response.
An agent architecture for understanding requests, selecting tools, retrieving structured data and producing analysis and reports.
Architecture and implementation components described
LangChain and MCP connect the agent to PostgreSQL and other data sources; local model serving is described through vLLM and a compatible model API.
RabbitMQ supports queued, asynchronous processing and component separation. An MCP server exposes analytical and system-integration tools.
Reports include indicators, trends, analysis and structured JSON output for dashboard visualizations.
Tool requests, responses and errors are logged; conversation history, concurrency controls and prevention of duplicate sensitive operations are described.
A model abstraction layer separates the application from a particular provider and standardizes requests and instructions.
Responses are aggregated for comparison using quality, relevance, completeness and consistency with the user's need.
The design supports additional models, tools and selection strategies and provides a path toward tool-using agent architectures.
Technical components
Provider abstractionParallel model requestsStructured response aggregation
The CV does not supply model rankings, evaluation weights, ground-truth labels or quantitative evidence that the selected response is always best.
Source: supplied AI portfolio · p. 6
05
Research infrastructure
Local language-model experimentation
Exploring model execution and reusable AI service components locally.
Local language-model environments and practical prototypes for application integration, multi-step processing and agent communication.
Environment setup and practical prototypes described
Local inferenceLanguage-model orchestrationConversation memoryTool integration
Described work
Ollama on Windows and LangChain / ChatOllama connect Python applications to local language models; CPU/GPU work distribution and response-speed factors are investigated.
LCEL processing chains, agent-to-agent communication, and MCP client/server architecture support multi-step tool and resource integration.
Conversation memory, streamed responses, retries, event/error logging and centralized error handling are described.
Gunicorn and RabbitMQ provide service execution and asynchronous message processing.
