AI Agent Infrastructure & Platform Engineering
AI Agent Infrastructure & Platform Engineering specialist building production agent runtimes, secure orchestration platforms, and agent memory. Creator of SynapsCLI (280K+ LOC Rust runtime, 51 stars), cut 21 production releases at Praxis AI, and directed Axon, an LLM inference engine built by an agent fleet that beat llama.cpp and Ollama in all 45 measured benchmark cells.
With 3+ years defending cloud infrastructure and now building agent-first architectures at production scale, I bridge security engineering with platform engineering β from Rust systems programming and ECS deployments to HMAC-authenticated capability gateways and live MongoDB migrations. Currently pursuing my Master's in Cybersecurity at NJIT.
When I'm not deploying to production at 2 AM, you'll find me building terminal UIs in Rust, red-teaming MCP protocols, or optimizing ONNX inference pipelines. π¦ + π€ + π
Lead TA for CS408/CS608 (Cryptography & Internet Security, 100+ students): delivered supplementary instruction on PKI trust chains, TLS handshake analysis and applied cryptography, and graded assignments and exams.
Specializing in cloud security, appsec, and cryptography. Currently serving as a Graduate Teaching Assistant.
Graduated with a minor in Computational Intelligence. Active in cybersecurity research and competitive CTF leadership.
Created and maintain an open-source agent runtime in Rust. v0.10.0 ("It's Daemon Time") made it a thin client over a shared session daemon: several terminals on one live session, sessions that survive restarts and resume, and secrets stripped before they reach the daemon. Also built subagent dispatch with zombie prevention, image input, and a benchmark suite (69MB @ 10 sessions vs 1.1GB for Claude Code). 33 releases across 5 channels (crates.io, GitHub, Homebrew, AUR, .deb). Ran 8 parallel agent reviewers on the 0.11 devβmain release, which led to 10 merged fix PRs. The runtime behind Praxis AI's agent VMs. 280K+ LOC, 5,000+ tests, 51 stars.
Directed a multi-agent engineering system that built an LLM inference engine from scratch in Rust and CUDA: 113K LOC, 130 merged PRs, no ggml or BLAS. Runs Qwen 2.5/3, Llama 3, Mistral and Gemma 3, plus Qwen3-MoE, gpt-oss and DeepSeek-V3-style MoE, on CPU and NVIDIA GPUs behind an OpenAI/Ollama-compatible API. Beat llama.cpp and Ollama in all 45 measured cells on Qwen2.5/Qwen3 dense models (0.5Bβ8B) across RTX 4090, RTX 5090 and H100 (up to 2.16x prefill, 1.70x decode vs llama.cpp), with md5-identical greedy output across batch paths. A foreman agent dispatches planner, builder and reviewer agents in isolated worktrees over a shared GPU queue; every PR must match main's output and pass benchmarks before merge.
Rust memory engine that keeps an agent's search index, signed memories and session handoff in one SQLite file. Merges VelociRAG (vector + BM25 + knowledge graph + metadata, fused with RRF) and Memkoshi (HMAC-signed memories). Consolidation strengthens what the agent uses, decays what it forgets, links related docs and prunes stale ones. Runs as a SynapsCLI extension that injects memories before each message and the last handoff at startup.
Open-source Rust TUI music client. 255 stars, 21 forks, 9 outside contributors, listed on Awesome Ratatui. Album-art-reactive color themes with cubic-eased cross-fade. 12K LOC, 200+ tests. Reviewed and merged 19 community PRs.
Research prototype synthesizing 7+ papers (Loopix, Vuvuzela, Pung): a 9-crate Rust workspace aiming to make "who-talks-to-whom" unanswerable to a global passive observer. ed25519/x25519 identities, ChaCha20-Poly1305 sealed boxes, a hand-rolled Double Ratchet, and encrypted headers over live Tor onion transport. 170 tests across 15 releases, #![forbid(unsafe_code)], key zeroization, and a published threat model.
Open-source security platform for Model Context Protocol (MCP). Dual-mode: static config scanning + real-time traffic interception via proxy, with response quarantine. 14 detection rules covering prompt injection, rug pulls, credential leaks, tool shadowing, and cross-tool attack chains. Caught the public MCP exploit PoCs it was tested on (Invariant Labs, Anthropic Git MCP, marmelab) and 83% of a 23-vector research corpus, with zero false positives on legitimate tools. 10K LOC, 242 tests.
Open-source 4-layer retrieval engine fusing vector search, BM25 keywords, knowledge graph traversal, and metadata filtering with RRF + cross-encoder reranking. ONNX-only β no PyTorch, no GPU, no API keys. Sub-200ms warm search. Published on PyPI, AUR, and the official MCP Registry. Now ported into Axel.
Open-source memory system for AI agents. Three-tier extraction (local regex, LLM API, pi RPC), staged review pipeline with HMAC-SHA256 integrity signing, and 4-layer semantic search via VelociRAG. Features pattern learning with access-weighted importance scoring, tiered boot context, staleness caveats, and bulk document ingestion. Cross-agent compatible β works with Claude Code, Gemini CLI, or any MCP agent. 321 tests. Now ported into Axel.
API-first security platform combining web vulnerability scanning (TLS, headers, XSS) with AWS cloud posture checks and CIS compliance mapping. Engineered modular recon pipeline with Docker.
Headless Linux server running 20+ Docker containers with Nginx reverse proxy, Pi-hole DNS, WireGuard VPN, and full observability (Prometheus, Grafana, Loki). Hosts a 12-agent AI orchestration system with semantic search and persistent memory.
Full-stack web app with Spotify OAuth integration enabling users to create personalized playlists merging top tracks from multiple artists.
Lightweight terminal-based audio player with URL/search playback and autoplay functionality using Python media libraries.
Led the web exploitation division across 50+ global CTF competitions.