I work at the intersection of LLM reasoning, MCP tooling, and SOC automation — building the security layer agentic systems need before they're trusted with real infrastructure, and building agents that do SOC work directly.

Supply-chain security scanner for MCP servers & Claude Code projects. Problem: MCP servers ship with zero standard vetting — tool-poisoning, command injection, over-broad permissions, and leaked secrets go straight into .claude/ directories. Stack: Python, zero runtime deps, SARIF 2.1.0 output for GitHub code scanning, CI-ready. Status: active, CI-ready

AI-powered vulnerability intelligence scanner. Problem: dependency vulnerability lists are noisy — a CVE hit doesn't tell you what's actually exploitable in your usage. Stack: Python · OSV.dev lookups across 7 ecosystems (Python, Node, Go, Maven, Ruby, NuGet, Rust) · LLM-backed exploitability + remediation analysis. Status: active

Local-first browser memory, accessible by Claude. Problem: Claude has no memory of what you've browsed — you re-explain context every session. Stack: TypeScript · browser extension + MCP server · fully local, no cloud sync. Status: shipped, v0.2.0

MCP server exposing defensive SecOps helpers: IOC extraction, defang/refang, hashing, password entropy, CIDR math, repo command-shadowing checks. Status: active

AI-powered SOC built entirely on Microsoft infrastructure — no Sentinel, no third-party SIEM. Problem: most SOC automation assumes a SIEM budget teams don't have. Stack: Copilot Studio · Power Automate · Microsoft Graph API · 5 coordinating agents (identity threat, phishing investigation, compliance/audit, SOC summary, orchestrator). Status: active

Open protocol for AI agent memory interoperability — like MCP, but for memory. Problem: every agent framework reinvents memory storage; nothing shares across vendors or sessions. Stack: HTTP-native spec · Memory Cell schema · access control · decay-ranked semantic search · active→stale→archived lifecycle. Status: spec stage, pre-PyPI

Local harness for the payment event failures that only show up in production: duplicate events, invalid signatures, malformed payloads, timeouts, upstream failures, partial writes, and out-of-order updates. Stack: Python - HMAC signature verification on the raw body - strict payload schema - idempotency by event ID plus payload hash - explicit order state machine that rejects stale events - bounded 3-attempt retry into a review queue - authorized replay with an audit timeline. Evidence: 18 deterministic tests cover the failure matrix end to end. No live provider is contacted. Status: active, public demo (https://github.com/glatinone/payment-webhook-repair-lab)

Streamable HTTP gateway in front of a small MCP tool registry, with a target smoke client that talks to it the same way a real client would. Stack: Node.js - per-request bearer auth (401/404/408/503 failure classes) - tenant isolation enforced at the gateway - request-bound expiring approvals for write tools - structured JSON logs with per-call request IDs and secret redaction - deterministic failure injection - runbook for each failure class. Evidence: 13 runtime tests cover auth, tenant boundary, approval binding, and each failure status. Status: active, public demo (https://github.com/glatinone/mcp-local-to-hosted-deployment-fix)

Bounded pre-release check for teams about to expose MCP tools to users or internal agents. Stack: Python - static rules for unsafe command declarations, excessive filesystem and network scope, secret-like values, and untrusted input reaching sensitive operations - secret-like values redacted before the report - local fixture server for tenant boundary, write approval, and quota checks - request-bound expiring write approvals - JSON and Markdown reports with rule ID, severity, evidence, and remediation - audit event per decision with reason code and sanitized input hash. Evidence: 10 deterministic tests cover the rule engine, tenant boundary, approval, quota, and report generation. The output states plainly that the preflight is not a penetration test. Status: active, public demo (https://github.com/glatinone/mcp-security-preflight)

Merged upstream:

Also tracking the MCP ecosystem via awesome-mcp-servers.

(Earlier school/web projects and one-off client sites not listed here remain visible in the full repository list.)