Hi, I'm Alex

Bachelor thesis platform, graded 10/10: two native detection engines plus an explainer, behind a single web interface. Argus classifies files, Aegis classifies URLs, and Athena turns each verdict into plain language. Every result sits side by side with the MetaDefender equivalent — Argus against its 21-engine multiscan, Aegis against its aggregated URL reputation.

A native multi-layer file scanning engine (Rust cdylib, shipped as argus.so/argus.dll) that fuses three layers into one verdict: hash and reputation lookups in LMDB, deterministic byte-pattern and fuzzy-similarity signatures, and an XGBoost model exported to ONNX and routed per file format by a selector. Format-aware PE and PDF parsers build the feature vector, and hyperlinks embedded inside documents are scored too. Standalone evaluation reaches F1 0.97–0.98 and ROC-AUC above 0.99 for both formats.

A native URL-classification engine (Rust cdylib) that reduces a URL to a canonical 107-feature lexical vector — structural, character, length, entropy, ratio and security signals — and scores it through an XGBoost model exported to ONNX, returning Clean / Suspicious / Infected without ever fetching the target page. Trained on nine public corpora deduplicated into a single set; F1 ≈ 0.96, ROC-AUC ≈ 0.99. Also embedded inside Argus to score hyperlinks found in documents.

The component that makes a verdict intelligible: a retrieval-augmented assistant answering questions over OPSWAT documentation and live scan results. Query rewriting contextualises follow-ups, hybrid dense + sparse retrieval over Qdrant fuses results with RRF, a cross-encoder reranks them, parent-child assembly builds the context, and a 24h Redis cache fronts generation — with Prometheus timings on every stage.

The model supply chain behind both engines: a ten-step DAG covering feature extraction, filtering, statistical profiling, Optuna tuning (a cold search then a warm-started refinement), XGBoost training, SHAP explainability, ONNX export and held-out evaluation — running either as an AWS SageMaker Pipeline or as local subprocesses. Feature parity is guaranteed by construction: the engine's own Rust extractor is vendored in and called through ctypes, so training sees exactly the features production computes.

A working RISC-V RV32I processor written from scratch in Verilog, with a three-stage fetch/decode/execute pipeline wrapped into a System-on-Chip — memory, serial port, timer and LEDs — targeting a Terasic DE2 Cyclone II FPGA and simulatable on any PC with Icarus Verilog. Documented for readers meeting processor design for the first time.

A modern, full-stack implementation of the classic Battleship game with real-time multiplayer capabilities, built using Test-Driven Development (TDD) principles.

BuildSync is a MERN-based construction management platform that streamlines workforce administration, attendance, leave, salaries, and client work requests through role-based dashboards for Admins, Employees, and Clients.

Uno Game is a digital version of the classic card game Uno, featuring real-time multiplayer gameplay, customizable rules, and an engaging user interface.