Building AI-powered products at Rectify — agentic assistants, real-time voice AI, MCP servers, and developer tooling. Self-taught since 2021, with 4+ years of professional full-stack experience.
I completed my HSC in 2020 — right in the middle of the COVID pandemic, a genuinely difficult time. Instead of continuing formal education, I chose to focus on building a career and taught myself software from 2021. What started as curiosity about web development grew into a deep focus on AI systems, developer tooling, and real-time applications.
Over 4 years of professional experience across internships, part-time, contract, and full-time roles. Today at Rectify, I work full-stack across 6 production repositories — shipping everything from LLM orchestration layers and agentic assistants to embeddable chat widgets and Chrome extensions. I care deeply about making AI systems that are reliable, cost-governed, and genuinely useful.
Outside of work, I stay current with the evolving AI ecosystem — the MCP protocol, voice AI, and multi-agent frameworks are things I've built with, not just read about.
▹Architected and shipped Quanta — an agentic AI assistant with 33 tools that autonomously controls the entire product dashboard, including realtime voice via Gemini Live (custom PCM16 AudioWorklet) and Twilio AI phone calls with µ-law/PCM codec bridging.
▹Built the Rectify MCP ecosystem: an npm-published MCP server with local Tree-Sitter code-graph intelligence and a backend HTTP MCP server exposing ~80 project-control tools to Claude/ChatGPT, secured by a hand-rolled PKCE OAuth implementation with RFC 9728 well-known metadata.
▹Designed the multi-provider LLM core — 6 providers (Claude, OpenAI, Gemini, DeepSeek, OpenRouter, Kimi) with priority fallback chains, RAG via MongoDB Atlas Vector Search, and a per-token cost-governance system covering 65+ models.
▹Developed AgentPulse: a full-stack monitoring platform for AI coding CLIs including the rectify-bridge npm CLI (reverse SSH tunneling, shell-injection-proof subprocess handling) and backend with Ed25519 device signing and LLM-based quality evaluation.
▹Built a production embeddable chat widget (AI-first + human handoff) with Socket.io real-time, rrweb session recording, markdown rendering, and media handling; paired with a 3-panel help-desk inbox and KB-grounded AI replies.
▹Shipped a Chrome Extension MV3 injecting voice-to-text, prompt enhancement, a 17-category code-snippet library, and web annotation into AI coding platforms using MediaRecorder, MutationObserver, and custom DOM injection.
▹Managed three distinct user interfaces, resolving bugs and implementing optimizations to improve performance.
▹Delivered responsive, user-centric features with consistent cross-device experiences.
ReactJavaScriptHTMLCSSResponsive Design
Projects
What I've built
Production systems shipped at Rectify — each with significant scope from design to deployment.
Quanta — AI Assistant
Private
33 agentic tools · Real-time voice · Twilio AI calls
Agentic AI assistant that autonomously controls the entire Rectify dashboard through 33 tools — generating reports, managing roadmaps, creating changelogs, and operating AgentPulse. Built end-to-end: LangChain ReAct agent loop on the backend with Socket.io token-streaming and a full-duplex voice interface on the frontend powered by Gemini Live and the Web Audio API. Also supports outbound Twilio AI phone calls with real-time µ-law/PCM audio codec bridging.
Published npm package implementing a Model Context Protocol server in two modes: a local repository intelligence layer (Tree-Sitter WASM parses TS/JS/Python into a persistent dependency graph with 7 tools for tracing, impact analysis, and search) and a cloud layer exposing Rectify project data. Additionally built the HTTP MCP server in the backend with ~80 tools across 8 domains, secured by a hand-rolled PKCE OAuth implementation (RFC 6749 + RFC 7636, no library) with full RFC 9728 well-known metadata.
Full-stack monitoring platform for AI coding CLI sessions (Claude Code, Codex). The npm bridge (rectify-bridge) creates a reverse SSH tunnel from the user's machine to the Rectify dashboard, parses local JSONL session files, and calculates real-time token costs — all with shell-injection-proof subprocess handling. The backend uses Ed25519 device signing for gateway pairing, LLM-based quality signal detection (refusal/frustration/laziness/context-loss), and trust scoring with 7-day decay. The frontend streams health data via SSE + Socket.io.
The central LLM platform powering all AI features at Rectify. Wraps 6 providers (Claude, OpenAI, Gemini, DeepSeek, OpenRouter, Kimi) behind a LangChain abstraction with a priority fallback chain, feature-key–to–provider routing, and instance caching. Pairs with a cost-governance system tracking per-token spend across 65+ models by feature, model, and period (daily/weekly/monthly/yearly) with 80%-threshold alerts and automatic entitlements enforcement backed by MongoDB Atlas Vector Search for RAG.
AI → human handoff · rrweb session recording · KB-grounded AI
Embeddable React chat widget (AI-first with Quanta → human agent handoff) and a companion 3-panel help-desk inbox. The widget handles full GFM markdown rendering, media uploads, reply threads, inline message editing, session transcripts, and real-time presence via Socket.io rooms. The inbox includes KB-grounded AI replies (vector similarity search), operating-hours–based human handoff, URL-serialized filters, dual pagination, and a 14-day trend / peak-hours dashboard.
Voice-to-text · Prompt enhancement · Snippets · Web annotation
Developer productivity extension injected into AI coding platforms (bolt.new, v0.dev, Replit, Lovable). Built four systems: voice-to-text via MediaRecorder API with real-time MutationObserver/ResizeObserver icon tracking and hold-shortcut detection; prompt enhancement with PRD/PRP management over Socket.io; a code snippet library with 17 categories and cursor-based pagination; and a web annotation system with CSS selector generation and session-storage badge overlays. Built with Vite + @crxjs and React 19.
