I'm a master's student in Computer Science at Blekinge Institute of Technology, Sweden. I build machine-learning and cloud projects: safety validation for automated driving, RAG systems, satellite-image segmentation, decision-support systems, time-series forecasting and Kubernetes microservices.
| Project | What it is | Result |
|---|---|---|
| afals | Falsifying the assumptions behind a safety-proven adaptive cruise controller with five search methods. | CMA-ES finds failing violations 8.8x smaller than random search; 300,000 nominal simulations without a collision. |
| risk-predictors | How much data a traffic-safety metric needs, on 139,410 real highway situations. | Time-to-collision metrics track a braking stress test (Spearman 0.44) where time headway does not (0.10); one recording is enough to pick the right metric family. |
| cutin-atlas | Parameterizing, clustering and efficiently simulating 2,322 real cut-ins. | A 4-parameter sigmoid keeps all 11 contact cut-ins; strata fitted on a pilot campaign cut simulation runs 4.7x. |
| synth-mix | How much generated weather imagery belongs in a detector's training set, and which images. | 31 YOLOv8 runs: low-drift generated images beat high-drift ones by 0.65 mAP points. |
| highway-lane-change-prediction | Predicting lane changes before they start on highD and exiD, with transfer between datasets. | On cars not yet moving sideways, boosting scores 0.68 against 0.34 for a one-line rule. |
| zod-fssl | Federated self-supervised learning on 100,000 real Zenseact frames split by country, and passing knowledge to a new car generation. | Distilling from the old model closes about two thirds of the gap to full labels with no labels. |
| driving-geometry | Data curation, a geometry-drift metric and camera calibration on real driving data. | Rare-scene lift up to 6.8 at a 5% budget (1.0 is random). |
| Project | What it is | Result |
|---|---|---|
| graphrag-multihop | Finds the second passage a multi-hop question needs by walking an entity graph and merging it with vector search. | All supporting paragraphs in the top 5 for 84.8% of HotpotQA questions vs 77.2% for vector search. |
| permission-aware-rag | Multi-tenant RAG with access control before ranking, a second check before the prompt and a tamper-evident audit log. | Retrieval matches an independent permission check on 40 random worlds; 93 tests. |
| semantic-cache-rag-router | Semantic answer cache with small/large model routing and a cost and latency dashboard. | Cache threshold set from 40,430 labelled Quora pairs. |
| trouble-report-investigator | Finds similar old bug reports and predicts the affected subsystem. | Hybrid search recall@10 0.830 vs 0.709 for BM25 on 14,037 Mozilla reports. |
| Project | What it is |
|---|---|
| tailwatch | Calibrated congestion-burst forecasts for a radio network, on two real 5G datasets, held out by group. |
| nfcast | Time-series foundation models vs classical methods on real Milan cellular traffic, with LoRA fine-tuning and a capacity-planning test. |
| multipath-eval | Comparing five multipath routing algorithms on generated and real network topologies. |
| evprior | Pre-trip and live remaining-energy prediction for electric cars with a Kalman filter. |
| fleet-memory | A shared road-condition map from many vehicles, with warnings that cite their evidence. |
| Project | What it is |
|---|---|
| marine-debris-segmentation | DeepSpectralUNet for finding marine debris in Sentinel-2 images (team project). |
| human-ai-text-classification | Audit of AI-text detectors: who gets wrongly flagged (non-native writers, new domains, evasion). |
| Project | What it is |
|---|---|
| askfirst | When a replenishment system should order on its own and when it should ask a planner, on Walmart M5 data. |
| depotpilot | A depot dispatch agent that repairs the day's plan and decides which decisions to send to a human planner. |
| Project | What it is |
|---|---|
| debris-fallout-risk-engine | Hackathon team project (EDTH, May 2026, 2nd prize): scores where debris from an intercepted drone or missile should not fall. |
| drone-selection-dss | Team project that ranks drones for small EU ports with AHP, TOPSIS and a RAG assistant. Live demo |
| netops | Kubernetes microservice app that flags anomalous node cards in real BlueGene/L supercomputer logs. |
chime-chat (Socket.IO, DynamoDB, S3) · svyj-chat (Firebase) · selenium-python-practice
Python, PyTorch, scikit-learn, FastAPI, SQL, JavaScript/Node.js, Docker, Kubernetes, PostgreSQL, AWS, Firebase.