Arjav Mehta

I build software and take it to market: compliance infrastructure, sports models, AI tools, and product launches. Finance concentration and data science minor at Northeastern.

mehta.arja@northeastern.edu GitHub LinkedIn Finance research
Company I founded2026 to present

LoxeAI: SOC 2 evidence your auditor can verify

Compliance tools show green checkmarks nobody can check. LoxeAI reads a startup's AWS account through a read-only role, maps what it finds to SOC 2 controls, and signs every piece of evidence so an auditor can confirm it independently. I built the scanner, the agent workflows, and the product site, and I lead sales and audit-firm partnerships.

15AWS services scanned across 6 regions
12SOC 2 controls mapped per scan
~40 hrsaudit prep saved per client
10+audit firm and practitioner partners
TypeScript, AWS (CloudFormation, IAM, Lambda, Athena), Cloudflare Workers, Supabase, LLM agents, MCP

loxeai.com

Product site. Run a free read-only scan of your AWS account and get a SOC 2 gap report.

Demo Center

Hands-on guided tours for startups, auditors, and partners, using a fictional company's data.

Pre-Audit Checklist

SOC 2 Type I readiness guide for 1 to 30 person AWS-native teams, with progress tracking.

AWS Evidence Layer

Open-source scanner: read-only IAM role via CloudFormation, 15 services, 6 regions, 12 controls.

LoxeAI MCP Server

Runs locally in Claude and other MCP clients. Scans infrastructure-as-code, maps findings to SOC 2 criteria, and drafts audit documents. Files never leave the machine.

Evidence Tracer Agent

The first MVP: maps CloudTrail events to SOC 2 controls through an IAM role.

Analyst intern projectSep 2026

NBA shot-make probability model for the OKC Thunder

Given player-tracking context for a shot (closest defender at release, shot clock, dribbles, shooter speed, location), estimate the chance it goes in. I blended LightGBM, XGBoost, and CatBoost on 39 features, recalibrated the blend, and added an empirical-Bayes estimate of each shooter's skill. Every modeling decision was tested by training on one season and scoring the other.

213,977test shots scored
0.6246out-of-season log-loss
0.6897base-rate log-loss, for comparison
39engineered features
Python, pandas, LightGBM, XGBoost, CatBoost, scikit-learn, pytest

The data belongs to the Thunder, so the code and writeup are shared on request.

Live tool for studentsOct 2026

The Voyager: co-op radar for Northeastern

Co-ops often appear on company career sites days before the school portal lists them. The Voyager polls public job boards on a schedule, classifies new postings, matches them to each student's resume with open-weight language models, and sends alerts. The poller respects robots.txt and backs off when a site asks it to. It runs on free tiers only.

9job-board platforms polled
$0monthly infrastructure cost
TypeScript, Cloudflare Workers, D1, Workers AI, GitHub Actions, Playwright
Open The Voyager Poller source
Personal AI system2026

Gideon: a voice assistant that runs on my laptop

A personal assistant with wake-word voice input, local speech recognition, and a model chain that falls back from cloud models to a local one. A small classifier decides when a reply should be brief and when it can have some personality, and it learns from my reactions.

TypeScript, Next.js, Python, Ollama, faster-whisper, openWakeWord

Private repository. Happy to walk through it.

Consumer launch2026

AMPHEX: a cognitive performance drink

Brand site and waitlist for a Loxe Labs beverage concept, built and launched end to end.

300+waitlist signups
Next.js, TypeScript, Tailwind, Supabase, Resend
amphex.pages.dev
Consumer launch2026

TheLingo: language learning as a ranked game

Two players answer the same speaking prompt and a blind judge picks the winner, so practice feels like a ladder match. The site includes a playable demo in five languages and a live waitlist.

45waitlist signups
Next.js, TypeScript
thelingo.xyz Source
Research proposal2026

LoxeRockets: mapping the ignorosphere

The 80 to 200 km layer of the atmosphere is too high for balloons and too low for satellites, so it is barely measured. This blueprint proposes flying recoverable sensor packages on sounding rockets to build a continuous dataset for reentry prediction, GPS disruption, and climate research. Submitted to NASA.

React, TypeScript, interactive blueprint
Read the blueprint
Side buildJun 2026

Wildlife Kingdom: a browser game

A quick game built with AI coding tools and shared on Reddit, where about 30 people played it.

JavaScript
Source
Finance2024 to present

Equity and credit research

Stock pitches, a credit analysis, and strategy case studies, collected in a separate portfolio.

financeportfolio.loxeai.com