Claude Code Specialist & Frontend Engineer
Claude Code specialist and frontend engineer with deep hands-on expertise in the full Anthropic ecosystem — Claude API, MCP protocol, and production Claude Code deployments. Founder of Albatross AI, where I've shipped pixel-accurate frontend builds, production MCP servers in Rust, and a Claude Code operating stack achieving 50-75% token reduction in daily use. I don't just use Claude Code — I build the infrastructure and interfaces around it.
01 / 05
Pixel-accurate frontend rebuild of an AI skin-analysis product
A multi-step flow — name/location entry, gallery/camera photo capture, live demographics results with confidence rings — rebuilt frame-for-frame from a real Figma spec and wired to live backend APIs, not mocked data.
A /code-review xhigh multi-agent pass caught 5 real bugs in one session, including a camera feed that never attached to the <video> element due to a ref-timing bug, and a localStorage unwrap bug that crashed the results page on real uploaded photos — found, fixed, and verified against the live API.
STATUS: Working, in progress — mobile pass pending
02 / 05
Hybrid AI-assisted security scanner
Semgrep performs real static-analysis detection as ground truth; Claude does the downstream triage — risk explanation, exploitability rating, fix suggestion — explicitly designed so Claude can never invent a finding Semgrep didn't flag.
Custom rules went from 1/4 to 4/4 detections on a planted-vulnerability fixture vs. community rule packs, zero false positives on the safe counterpart. A second 'AI reasoning' pass catches business-logic bugs invisible to rule-based scanning — run against its own code, it caught a real bug: the FastAPI server was bound to 0.0.0.0 instead of localhost.
STATUS: Verified end-to-end locally via real HTTP calls and MCP round-trips; not yet deployed
03 / 05
RAG job copilot + a 26-panel live infrastructure dashboard
An AI Job-Hunter Copilot (LangGraph agent, hybrid BM25 + TF-IDF retrieval) that ranks leads and drafts cover letters, paired with Omni Dashboard — a live monitor surfacing real-time health across every service in the stack, plus a multi-platform social posting panel.
Caught and fixed a real retrieval bug live: vague queries like 'rank my top 3 leads' shared no vocabulary with the lead documents, so lexical retrieval missed them — fixed by bypassing similarity search for that intent. Omni Dashboard uses an honest three-state status model (green/orange/red) instead of binary healthy/broken, so free-tier API limits stay visible rather than hidden.
STATUS: Omni Dashboard is live; RAG system verified locally
Live Dashboard →04 / 05
Production Rust MCP server for agentic infrastructure
Gives Claude Code live tools to check Render + Vercel deployment status, query Buffer content schedules, read Firestore leads, and send email via Gmail — a real JSON-RPC 2.0 MCP server over stdio.
An earlier version faked its tool responses with hardcoded data over a non-standard protocol. The current version is a genuine rebuild: real JSON-RPC 2.0, live API integration, a security-hardening pass moving API tokens to headers, and passing unit tests.
STATUS: Working, real local stdio MCP server — the automation backbone for job-lead-discovery
05 / 05
Automated YC Jobs board scraper and lead digest
Scrapes the YC Jobs board, parses each listing, filters to US-only roles, and scores leads by keyword match against a target-role persona — then emails a ranked digest.
Scoring is an honest additive keyword-count heuristic, not ML-scored — deliberately simple and legible rather than oversold. The email step calls into agentic-rust-mcp's send_gmail tool, though that cross-repo path is currently stale from a repo reorg and needs a fix to run end-to-end.
STATUS: Scraper/parser/filter pipeline verified against the live YC Jobs page; email integration needs a path fix