Student · Founder · Engineer

Xavion Mirchandani

I build AI products and the systems around them.

I'm a student at Blair Academy who built Xavion AI, a multi-model AI platform. This summer I interned as a software engineer at Preface in Singapore, where I shipped production AI agents and built the company's internal CRM.

View projects About me

01 — Experience

Where I work

Jun 2026 – Aug 2026
Singapore
Preface — Software Engineering Intern
Sole engineer on the Singapore team of Preface, a global technology-education company — built agentic automation and internal tooling that runs in production for real client outreach.
- Built and ran a production multi-agent pipeline in Python that automates client outreach end-to-end: an outreach agent that drafts personalized emails with AI-researched company backgrounds, a scheduling agent that classifies reply threads and flags follow-ups, and a calendar-sync agent that reconciles Google Calendar with color-coded client statuses.
- Designed and shipped Preface 2559, the company's internal CRM platform — a Prospect → Won pipeline with a drag-and-drop board, per-business timelines, and a human confirm queue that gates every outward-facing agent action. Deployed at preface.sg.
- Hardened the platform for production: database migration baselines with CI, row-level security, a durable action executor with idempotency keys and full audit trails, and an agent evaluation suite of 40 fixtures including prompt-injection attacks.
- Python
- TypeScript
- Next.js
- Supabase
- PostgreSQL
- Gemini API
- Google Workspace APIs
- Vercel
2025 – 2026
Singapore
Xavion AI — Founder & Developer
Founded and solo-built a multi-model AI platform that brings 30+ models from every major provider into one interface.
- Designed a unified abstraction layer over multiple AI provider APIs (OpenAI, Anthropic, Google, and others), normalizing authentication, request formats, and response structures.
- Implemented real-time streaming with Server-Sent Events and side-by-side conversations with up to six models at once, plus image generation and prompt refinement.
- Rebuilt the entire interface from scratch after the first version fragmented into separate pages — one product decision that shaped how I think about design.
- Next.js
- React
- TypeScript
- SSE
- Provider APIs

02 — Projects

Featured projects

Xavion AI

AI platform · Founder & solo developer

A multi-model AI platform that brings 30+ models together in one place. Chat with up to six models side by side, compare their responses, generate images, and refine prompts — all from a single interface.

Next.js
TypeScript
SSE streaming
30+ models

Visit Xavion AI Write-up

Preface 2559

Internal CRM platform · Preface internship

An agent-first CRM built for Preface during my internship. Businesses move through a Prospect → Won pipeline on a drag-and-drop board, every email, meeting, and stage change lands on an append-only timeline, and AI agents propose actions that humans approve from a confirm queue. Runs in production with row-level security, audited action execution, and CI.

Next.js
TypeScript
Supabase
PostgreSQL
Gemini API

Visit the CRM

Home Lab Server

Home server · Personal project

A home server I designed and built from parts: 360TB of raw storage across 18 drives in three RAIDZ2 vdevs, 192GB of ECC RAM, and GPU passthrough for a Windows VM. Runs TrueNAS SCALE with Plex, game servers, and a stack of self-hosted services.

TrueNAS SCALE
ZFS
360TB storage
Self-hosted

Visit the server homepage

Personal Investment Portfolio

Stock portfolio · Since 2022

A portfolio I've managed myself since 2022, built around a core of long-horizon mega-cap tech positions and a satellite of smaller positions I actively rebalance. The tracking dashboard fetches live market data to compute real growth — roughly 90% to date — with every trade and its reasoning on record.

Investing
Portfolio management
Live market data

View the dashboard

Big Two

Card game AI · Personal project

A complete Big Two engine and two independent AI opponents. The first is a determinized Monte-Carlo bot that computes exact hypergeometric beat probabilities and decomposes a hand with set-partition dynamic programming; the second is a DouZero-style deep Monte-Carlo network trained by self-play. The rollout hot loop is hand-written C, bridged through ctypes with a pure-Python fallback.

Python
C
Monte Carlo
Deep RL

Play Big Two

Mahjong Analyzer

Hand analysis engine · Personal project

Give it a 14-tile hand and it ranks which scoring hands are closest, accounting for jokers, exposed tiles, and everything already visible on the field. It precomputes roughly 21 million target configurations into a memory-mapped binary and scans them in a hand-written cross-platform C engine for sub-second evaluation — bound into both ctypes and Android JNI, so the web app and the phone app run identical code.

C
Python
React Native

Google PlayApp Store

All projects

03 — Writing

Latest writing

Mar 15, 20262 min read

How I Built Xavion AI — A Multi-Model AI Platform

The story behind building an AI comparison platform with 30+ models as a solo developer: learning the stack from zero, streaming architecture, and the rebuild that taught me the most.

In the queue

Shipping AI Agents to Production at PrefaceIn progress
Building My Homelab — A 360TB Home ServerIn progress
Building a Pinball Game in ScratchIn progress

All writing

04 — Contact