Open Source, projects, and experiments.

I designed the two-pass architecture (Understand first, Secure second), modular Python CLI toolkit for token-optimized autonomous agent authoring, deterministic web ingestion, and static documentation site deployed with Astro and Starlight. Covers trust boundaries, tool authorization, prompt injection, and secure reference architectures.

I worked on the context-routing, VitaeGraph structured memory, and native MCP server integration for a portable framework with eight focused skills and adapters for 11 agent environments. The zero-external-dependency MCP server exposes four resources, three tools, and four reusable prompts while preserving evidence boundaries and deterministic context retrieval.

I developed a fraud-detection agent stack with LangChain and Langfuse, then tuned recursion depth, token use, and sampling settings to balance accuracy, cost, and latency under a tight contest budget.

I trained and evaluated a multi-model detection pipeline across network flows, malware API traces, unseen attacks, and shell-command tactic attribution. Selected results include 0.82 test macro F1 for the strongest malware models and 92% anomaly recall, alongside shortcut-learning and minority-class analyses.

I built evidence chains across source, bytecode, verifier logs, translated instructions, and runtime traces. The analysis confirmed one taint-tracking bypass and one bounded 54-byte intra-packet overflow; neither established kernel RCE or privilege escalation.

I built an investigation workflow where players inspect five evidence types and assess supported and hallucinated AI claims against chain-of-custody ground truth. An evidence-review gate prevents verdicts before raw artifacts are inspected.

I designed a multithreaded Python CLI that integrates Hybrid Analysis, VirusTotal, and ANY.RUN, then mapped 12 Equivocal Software Behaviours from 60 MITRE ATT&CK Enterprise techniques. The thesis analyzed 36 SourceForge goodware binaries and was published on Zenodo.

I built preprocessing and evaluation pipelines for seven intent categories using classical ML, clustering, and BERT representations. Supervised and frozen-BERT classifiers reached about 0.99 weighted F1 on the dataset while remaining weaker on rare behavior.

I implemented core cryptographic primitives and protocol exercises in C and Python, then stress-tested them through cryptanalysis and CTF-style break/fix challenges. It bridges formal concepts with hands-on attacker thinking.

I implemented custom QEMU initialization for memory, clocks, NVIC, LPUART, and PIT timers, then ported FreeRTOS and built event, monitoring, and alert tasks. The MPU study documents an adaptation route rather than claiming a completed MPU port.

I designed and compared IPSec and TLS VPN topologies in GNS3 using Cisco routers and Dockerized endpoints, then verified IKE/ESP security associations, TLS handshakes, NAT traversal, SecureNAT assignment, and protocol trade-offs with Wireshark and SoftEther logs.

I processed Android GNSS measurements for pseudorange and spoofing analysis, then ran WiFi goodput experiments with iperf, tcpdump, and Wireshark across Ethernet, WiFi, and mixed scenarios with repeated min/max/average/std-dev reporting.

I developed a security-first full-stack forum with server-enforced role and ownership checks, session-based authentication, TOTP-gated admin moderation, salted scrypt password verification, parameterized SQLite queries, and anonymous comments.

I co-built Budget HQ, a fintech-focused web platform created during the IEEE-HKN Global Hackathon where we secured 1st place worldwide. The delivery balanced rapid product iteration with secure full-stack engineering.