ML Researcher at Iquadrat
Dimitrios Tsoukalas
Machine Learning Engineer & Researcher
I work with computer vision, NLP, and reinforcement learning. Right now I'm researching how large language models can help manage 6G networks.
Projects
RouteCause — BGP Incident Investigator
A citation-grounded BGP incident investigation tool. Deterministic analyzers detect the anomaly first; an agentic LLM then explains it, citing evidence and abstaining rather than guessing when the evidence doesn't support a claim.
- Detects BGP hijacks, route leaks, AS-path loops, and MOAS events with deterministic analyzers, not the LLM
- Hybrid retrieval (BM25 + dense embeddings + Reciprocal Rank Fusion) over RFCs and routing data, with NLI-based citation verification
- Searches for contradictory evidence and can abstain instead of asserting an unsupported claim
- FastAPI REST API with Docker deployment, OpenTelemetry tracing, and an offline evaluation harness backed by 123 tests
5G RAN Root Cause Analysis Assistant
A RAG-powered assistant that finds the root cause of 5G network faults. Ask a question in plain English and get a clear, source-cited answer in seconds. It runs fully on your own machine, with no OpenAI API and no cloud costs.
- Classifies faults as RSRP drop, handover failure, or throughput collapse
- Retrieves evidence from 188k real 5G KPI measurements across 83 CSV files
- Grounds answers in 15 structured NOC incident reports
- Recommends concrete remediation steps per fault type
On-Device Finance Optimizer
A privacy-focused app that extracts and categorizes transactions from receipts and bank statements using locally hosted LLMs, so financial data never leaves the machine.
- Parses PDFs and images via PDFPlumber and OCR (Tesseract/EasyOCR), normalizing results into categorized transactions
- Runs entirely offline through Ollama, using Mistral 7B, Llama 3.2, and Qwen 2.5 for extraction and categorization
- Streamlit interface with batch processing and CSV export