Full-stack Engineer building production systems end to end. I combine high-performance frontend engineering with backend architecture that holds under load, including chat infrastructure, background workers, AI pipelines, caching layers, and event-driven workflows.
Developed the complete real-time AI infrastructure for an emotional support agent that leverages multi-modal inputs to detect risk, evaluate mood, and deliver personalized counseling.
- Architected a high-performance multi-model pipeline using Gemini 2.5 Flash for sub-second risk classification (0–4 scale) and Gemini 2.5 Pro for deep, clinical-grade session synthesis.
- Engineered a Universal OpenAI-Compatible Hook (/v1/chat/completions) that transforms the backend into a pluggable "Brain" for external voice agents like ElevenLabs, managing real-time SSE streaming and session mapping.
- Built a Contextual Memory Engine utilizing text-embedding-005 and PostgreSQL (Drizzle) to maintain long-term user history, ensuring every interaction is grounded in past "anchors."
- Developed deterministic safety guardrails via a custom Intervention Arbiter, enforcing Zod-validated schemas and risk-level gates to provide safe, empathetic mirroring without clinical diagnosis.
- Optimized for Voice-First Latency on Google Cloud Run, orchestrating concurrent "First-Response" and "Counselor" streams to keep interaction fluid and human-centric.
Stack: Hono, React, Google Cloud Run, PostgreSQL, Drizzle ORM, Vertex AI (Gemini Flash, Pro, Embeddings), ElevenLabs, Pulumi
Architected a high-performance, serverless AI companion that transforms raw League of Legends match data into actionable coaching insights using AWS Bedrock and React 19. Delivered the app in a 4-week sprint with a 2-person team.
- Managed the full-stack Turbo Monorepo on AWS SST (Lambda/Node.js 22), achieving <15ms latency for cached operations.
- Built a robust JSON-only architecture with Amazon Nova Micro, utilizing strict schema enforcement and token-optimized prompts to convert complex match logs into machine-readable coaching data.
- Implemented "Graceful Degradation" logic where invalid AI outputs trigger deterministic fallbacks, ensuring 100% UI uptime even during LLM failures.
- Developed a Global Rate Gate middleware to orchestrate Riot API consumption, managing concurrency and quotas across distributed services.
- Implemented on-the-fly social card generation using server-side rendering stored in S3.
Stack: React, TypeScript, Hono, AWS Lambda (SST), Amazon Bedrock (Nova), TanStack Query, Tailwind, ShadCN
Built the full backend and AI orchestration layer for an agentic system that converts raw Asana social-media tasks into validated schedules and repost recommendations.
- Designed the end-to-end architecture, mapping how CSV/XLSX data moves from frontend parsing into backend normalization and finally into the Orchestrate agent.
- Implemented the backend with Hono.
- Integrated IBM watsonx Orchestrate using a custom JSON contract that drives multi-step agent reasoning for captions, validation, post-type classification, and scheduling.
- Built the agent’s knowledge base, combining brand guidelines, product features, rule logic, historical posting data, and a formal prompt spec for deterministic output.
- Delivered the entire reasoning pipeline, ensuring stable, machine-readable JSON for the frontend preview experience.
Stack: React, Hono, Cloudflare Workers, TypeScript, IBM watsonx Orchestrate
Core engineer for a counseling platform connecting clients with psychologists.
- Built the backend with NestJS and PostgreSQL (Drizzle ORM).
- Added real-time chat with presence, file sharing, and unread tracking using Ably.
- Integrated Zoom automation, notifications, and background jobs with Bull + Redis.
- Containerized and deployed the app on GCP Cloud Run using Docker.
- Added retry logic and monitoring to background tasks.
Stack: NestJS, PostgreSQL (Drizzle), Redis, BullMQ, Ably, React
Worked on the frontend for an AI-powered translator that converts Malay Sign Language into text in real time.
- Built a responsive gesture recognition interface using TensorFlow.js and MediaPipe.
- Integrated PAN AI LLMs for contextual text generation and improved translation accuracy.
- Deployed via GCP Cloud Run for seamless browser-based interaction.
Stack: React, TensorFlow.js, MediaPipe, GCP, PAN AI LLMs
