PhD student at Università della Svizzera italiana (USI), Lugano. I work on agentic AI, retrieval-augmented generation (RAG) and caching for LLM systems.
I build the layer between LLM agents and the data they reason over. My goal is to make agents faster, cheaper and more reliable by reusing what earlier runs already learned, instead of re-retrieving and re-reasoning from scratch every time.
- KAIROS (Knowledge Architecture for Incremental Reasoning Over Data Sources): middleware between LLM agents and evolving knowledge graphs. Agents query Neo4j through an MCP server, and a semantic cache stores validated Cypher queries, results and distilled facts (never final answers), so similar questions skip rediscovery.
- An agentic graph-QA benchmark built on Neo4j's GRAPE: LangGraph agents with MCP tools, an LLM-as-judge, concurrent evaluation and per-tool-call instrumentation. Runs on cloud LLMs (OpenRouter) and local ones (llama.cpp).
- Exploring: approximate and semantic caching for RAG, KV-cache reuse, vector search (ANN/HNSW), and efficient serving of retrieval-augmented systems.
Agentic AI · Retrieval-Augmented Generation · LLM caching · Model Context Protocol · Knowledge graphs · Vector databases · Reinforcement learning · Distributed systems
| Project | What it is |
|---|---|
| ai-council | Ask Claude, ChatGPT and Gemini side by side, then let them review and improve each other's answers over several rounds. Runs locally on your existing subscriptions, with lean prompts and token and limit tracking. |
| Task-offloading | M.Sc. thesis: a deep RL policy that decides whether metaverse tasks run on the device, the edge or the cloud, balancing soft/hard deadlines against device energy. Compares a central cloud agent with one agent per edge server. Includes an English summary report. |
| Lifeware-Challenge | Neural Wave AI Hackathon (team NAMA): reverse-engineered a legacy insurance system into deterministic Python. Generic "rule detective" tools (shallow decision trees, change-point detection, residual analysis and convention search) surfaced the hidden rules. |
| GATN | Graph attention transformer for multi-label retinal disease classification: EfficientNet features plus a label graph built from biomedical text embeddings. |
| Deep-Learning | PyTorch coursework, including GAN-BERT for machine-generated text detection (SemEval-2024), which beats a BERT baseline at every label fraction. |
| RL | Deep RL: custom Gymnasium environments, n-step SARSA and Q-learning, DQN vs. Double DQN. |
| game-theory-surgical-allocation | Deferred-acceptance matching of surgery requests to hospitals under Poisson arrivals. |
| Data-Network | OMNeT++ simulations of data-link, MAC and OSPF protocols, plus Linux networking. |
Earlier work: Adaptive Noise Reduction (LMS, NLMS and RLS in MATLAB) · 2×2 Rubik's Solver (C++ graph search) · Course Media Player (PyQt5) · Advanced Programming (C++17 and Python)
- PhD, Università della Svizzera italiana (USI), Lugano: agentic AI, RAG and caching (current)
- M.Sc. in Electrical Engineering (Communication Systems), Sharif University of Technology, 2025. Thesis: Design a Task Offloading Policy for Computational Tasks in Metaverse, a deep RL policy for end–edge–cloud systems comparing single-agent (cloud) and multi-agent (edge) designs
- B.Sc. in Electrical Engineering, Amirkabir University of Technology (Tehran Polytechnic), 2021
- Neural Wave AI Hackathon (Lugano), team NAMA: reverse-engineered a legacy insurance system's calculation rules into deterministic Python, matching the legacy outputs on 99.7–100 % of the training cases