San Francisco and Miami

Rayan Arya

I'm a Data Science and Artificial Intelligence student at the University of Miami, where I'm part of the Foote Fellows Honors Program and pursuing minors in Mathematics, Quantitative Economics, and Business Technology.

I like building things that are both technically interesting and genuinely useful. Some of my projects focus on consumer software, including a voice-first AI teammate, a calendar assistant that schedules through plain language, and a fully local AI notetaker. Others are more quantitative, like pricing options, estimating fair odds, and testing whether a model can actually outperform the market.

I've also worked on research involving drone imagery and digital pathology, where I build machine-learning pipelines and evaluation systems that measure not only what a model predicts, but how much confidence we should place in it.

rayanarya25@gmail.com GitHub LinkedIn

StudyingB.S. Data Science and Artificial Intelligence, 2029
MinorsMathematics, Quantitative Economics, Business Technology
BasedSan Francisco and Miami
01Research

Work where the question came first and the software existed to answer it.

Tree detection from drone imagery

Frost Institute for Data Science and Computing

The hard part is deciding what counts as correct, and that decision came from a measurement.

A two stage pipeline that replaces manual tree marking: detect individual trees from drone orthomosaic imagery first, then classify species. It runs across the University of Miami campus and Big Cypress National Preserve.

Ground truth is 10,659 botanical inventory points. The prior machine generated polygons missed about 58% of the confirmed inventory trees and merged adjacent crowns in roughly 640 cases, so the labelling strategy trusts the points and discards the polygons.

DeepForestYOLOPyTorchRemote sensing

Options analytics

Python, pandas, NumPy

Pricing theory turned into something you can actually size a trade with.

At the money implied volatility solved with Newton Raphson, a full Greeks surface, realized volatility measured against implied across the term structure, straddle based pricing with confidence intervals, and a probability cone for sizing a position.

The comparison is the point. Implied against realized is what tells you whether options are expensive or only look expensive.

Newton RaphsonGreeksStreamlit

BetAI

Fair odds models for NFL and NBA

The output is a probability, not a pick. The bet is whatever the gap to the market says it is.

Fair odds models for NFL and NBA games, built from public ESPN box scores and Sleeper injury data. It pulls completed games, builds rolling pre game features and home minus away differentials, and trains a calibrated classifier that estimates a true win probability rather than picking a side.

For upcoming games it layers an injury based log odds adjustment on top, using Sleeper's injury tags because they are more reliable than ESPN's. The NBA pipeline goes further and models player props, points, rebounds, assists and threes, and writes out over under recommendations.

ESPN APISleeper APIscikit-learnProbability calibrationpandas

Intraday signal study

SPY, five minute bars

A negative result, reported as one.

A direction prediction study on SPY built to be methodologically honest rather than impressive. Five minute bars from yfinance, features across lagged returns, RSI, MACD and Bollinger bands, and a random forest trained on a chronological split so nothing leaks backwards from the future.

It reached 57.3% on 103 test samples, which is not statistically distinguishable from a coin flip, and the writeup says exactly that. The repository is the pipeline and the honesty about its limits, not a trading strategy.

yfinanceRSI, MACD, BollingerRandom forestpandas
02Products

Things with users, or on their way to having them.

Alfred

Co-founder, engineering and product

A shipped desktop product on two platforms, and the team around it.

A voice first, screen aware AI teammate. It watches the window you are actually working in and answers about what is on it, instead of asking you to paste context into a chat box.

I work across the whole stack: large parts of the macOS client in Swift 6 on ScreenCaptureKit and WhisperKit, the Windows port, and the Cloudflare Worker that fronts the Anthropic API. Around 400 commits so far, from beta auth and rate limiting to DMG packaging and release engineering.

Swift 6ScreenCaptureKitWhisperKitCloudflare WorkersAnthropic API

TimelyCal

Full stack engineer, web and desktop

Full stack ownership, from the frontend to the agent to the database.

A chat based calendar assistant that turns plain language into scheduling across Google and Apple calendars. I built the Next.js frontend, the Electron desktop client, the Gemini agent that does the scheduling, and most of the Hono, tRPC, and Postgres backend behind it.

It does conflict aware booking across both calendars at once and keeps a persistent memory of your preferences, so it stops asking you the same question every week.

Next.jsElectronHono and tRPCPostgres and DrizzleBetter-AuthTurborepoGemini API

Doop

Native macOS, everything on device

Frontier model behaviour on hardware you already own.

A meeting notetaker that never phones home. Audio, transcripts, embeddings and summaries all stay on the machine, which is what makes it usable for conversations sitting under an NDA or a client contract.

It captures in person conversations and Zoom, Meet or Teams calls through Core Audio process taps, with no third party audio driver to install. WhisperKit transcribes, SpeakerKit separates who said what, and a local Llama model summarizes.

Swift 6Core Audio tapsWhisperKitSpeakerKitOllamaGRDB
03Hackathons

Built in a weekend, which is usually the fastest way to find out whether an idea works.

Hindsight

Fourth place, YC GStack and GBrain hackathon

Memory teaches an AI your facts. This teaches it your patterns of being wrong.

A calibration layer for AI advisors. It pulls the testable claims out of what someone has written, grades them against what actually happened since, and builds a profile of where their judgment runs hot.

Five GBrain skills with an Express server behind them and a browser sidebar on top. Run over 228 Paul Graham essays it scored a 67% hit rate and surfaced the decision domains where overconfidence shows up.

TypeScriptGBrain skillsExpressCalibrationMerged upstream

Goodhart

Finalist, YC HUD Frontier and RSI RL Environments hackathon

Finding the way to cheat the grader before a policy finds it.

An automated red teaming system for reinforcement learning graders. The graders people actually use are weak and easy to game, so they quietly teach a model to cheat rather than to solve.

Red team agents hunt for solutions that pass the grader without being correct, a held out oracle rules on whether a breach is a real flaw, and a green team agent hardens the grader. A regression gate makes sure each patch blocks the cheat without breaking the correct solutions.

Reinforcement learningReward hackingAutomated red teamingClaude API

Envoy

YC Voice Agents hackathon

It learns from the attack that worked, and it learns during the call.

A voice first booking agent that calls a restaurant, negotiates a time, and confirms the reservation with nobody in the loop, wired into Google Calendar and Gmail so the result lands where it belongs.

When a social engineering attempt actually works, it does not just log it. A separate Claude judge reads the transcript and flags the leak, a patcher writes a guardrail against the class of attack rather than the exact wording, and the rule hot reloads into the live pipeline before the agent's next turn.

PipecatNVIDIA NemotronTwilioSelf healing guardrailsGoogle APIs
04Experience

The roles behind the work above, in order.

April 2026
to now

Alfred Co-Founder, Engineering and Product

Technical co-founder of a voice first, screen aware AI teammate. I work across the macOS client, the Windows port and the backend, and lead a three person founding team while we work toward a pre-seed raise.

May 2026
to now

TimelyCal Full Stack Engineer

Lead web and desktop for a chat based AI calendar assistant, from prototype to production, alongside two non technical founders and one other engineer.

Summer 2026

Intracellular Technologies Data Science Intern

Two Java and QuPath extensions for a digital pathology pipeline. The first exports pathologist annotations from Aperio whole slide images into ImageScope compatible XML for handoff to scanning partners. The second pushes annotation GeoJSON, provenance metadata attached, to AWS S3 and DynamoDB. I owned both end to end, down to the coordinate space mapping and cross platform validation.

Summer 2026

Frost Institute for Data Science and Computing Research Intern

Machine learning pipelines for individual tree detection and classification from drone orthomosaic imagery, covering the data pipeline, model training, and evaluation methodology.

August 2025
to now

TAMID Group at Miami Quantitative Analyst

Equity analysis and trading strategies in Python with pandas and yfinance, plus the pipeline that validates, versions, and releases the chapter's quantitative research. Repository

Summer 2024

UCLA Visual Intelligence Lab Summer Research Intern

Replicated Deep Compositional Question Answering with Neural Module Networks by Andreas et al. in PyTorch under a graduate mentor, modernizing the original implementation so the results could actually be reproduced. Repository

05Stack
Languages

Python, TypeScript, JavaScript, Java, Swift, R, SQL

Models

PyTorch, DeepForest, YOLO, Whisper, Ollama, the Claude and Gemini APIs, agents and tool use, eval design

Product

Next.js, React, Hono, tRPC, Postgres, Electron, Streamlit, AWS with S3 and DynamoDB, Gradle, Git, pandas, NumPy

06Background and life outside it
Education and awards

Foote Fellows Honors Program at the University of Miami, finishing a B.S. in Data Science and Artificial Intelligence in 2029, with minors in Mathematics, Quantitative Economics, and Business Technology. Provost's Honor Roll in 2025. Before Miami I was at The Harker School in San Jose.

I'm VP of Partnerships and Outreach for Scholars of Finance. At the UM TargetCanes case competition I placed in the top ten out of more than fifty, competing on my own against teams, on a case built around local, on device AI.

Outside of work

Football and basketball, which is where the prediction models came from in the first place, and poker for more or less the same reason.

Surfing and hiking whenever the coast or the trail is close enough. Photography. Bhangra, which is usually the one that surprises people. And music, which mostly means going to as many concerts as I can get to.

Let's build something

Always glad to talk about agents, evals, research, or early stage engineering. Email is the fastest way to reach me.

rayanarya25@gmail.com GitHub LinkedIn