Aryan Malik

Aryan Malik is graduating from Rutgers University–New Brunswick (Honors College) in December 2026 with a B.S. in mathematics, computer science, and economics, and is looking for full-time work in San Francisco from January to May 2027. He is working toward becoming a quant engineer who owns research, development, and trading. He built Trade the Knight, the Rust exchange behind Rutgers’ first live algorithmic-trading competition; Osmo, a self-hosted personal AI agent; Aequus, on-chain markets on any real-world quantity; and this site, where people trade prediction markets on his real health and activity data. He is the math, CS, and data science chair of the Rutgers Quantitative Finance Club, works as a software and data intern at LuminArx Capital in New York, has won first place at the Rutgers/KPMG Ideation Challenge, the SQC Quant Challenge, and the BITS/Barclays Datathon, scored 1600 on the SAT, and holds a third-degree black belt.

Work

• Software and Data Intern, LuminArx Capital Management (Jul 2026 – present) — Built Lumin, a proactive company agent for the investment and operations teams, and is moving the firm’s data pipelines to Azure.
  • Lumin is a proactive company agent that watches firm email and Teams traffic and flags problems for the investment and ops teams. In its first week it served 30 users over 200+ messages and caught about 10 real issues on its own.
  • Lumin’s knowledge base is Markdown in SharePoint behind an MCP server with keyword and semantic search, updated autonomously from documents, email, and Lumin’s own conversations.
  • Moving about 10 data sources (about 10–15 GB a day) from ad hoc SFTP scripts and consultant-run jobs to firm-owned Azure pipelines.
  • Consolidates positions and deal-room data into a single store that agents can query.
  • Builds dashboards for exposures, returns, NAV, and reconciliation.
• Researcher and Developer, Independent (May – Aug 2026) — A Python/Rust mid-frequency trading stack, an autonomous research loop, and an execution-cost model, built independently.
  • Python/Rust mid-frequency trading stack: a Rust Thetadata FPSS streaming client (600+ symbols), risk daemons (pre-trade limits, kill switch, stale-data halts), and a Rust gRPC execution service routing to Interactive Brokers (IBKR, paper trading).
  • Autonomous research loop: researcher, trader, and risk-officer agents seed hypotheses from arXiv, experiment in a sandboxed Docker notebook, sweep backtests, and gate ideas on Deflated Sharpe and train/test gaps, with memory of past trials. A human makes the final call.
  • Execution-cost model trained on Databento MBP-10 data: about 1.0 bp out-of-sample error on $10k slippage and 1.9 bp on spread.
• Quant Researcher/Developer, Gradient Trading (Jan – Jul 2025) — Quant research and infrastructure at an early-stage quantamental fund.
  • Built a Python level-2 backtester that models spread and slippage; simulations landed within 50 bps of live results.
  • Trained a PyTorch time-series attention model to reconstruct level-2 order-book depth from level-1 data.
  • Built ingestion pipelines for Bloomberg, USPTO, and SEC filings, extracting about 2 TB.
  • Ran containerized Docker/Airflow pipelines processing 3 GB a day, monitored with Grafana and Prometheus.
• Data Science Extern, New Jersey Big Data Alliance (May – Aug 2024) — NLP for credit transfer, and a statewide data science curriculum.
  • Built an NLP matcher on a local quantized embedding model that mapped 1,138 course equivalencies from 2-year to 4-year degrees, 100% accurate on a held-out eval set plus spot checks.
  • Designed a statewide data science curriculum, benchmarked against UC Berkeley and UT Austin, that was adopted across 9 colleges.
• Math, CS & Data Science Chair, Rutgers Quantitative Finance Club (May 2024 – present) — Runs the club’s research teams, competitions, and infrastructure.
  • Directed 9 research teams replicating foundational quant papers, including Heston and SABR.
  • Trained forward-backward neural SDEs with PPO on Heston-model data and showed they would need 100+ years of tick data to converge.
• Research Lead, Rutgers Economics Lab (Sep 2024 – present) — Leads applied research for public agencies.
  • Led a 6-person team quantifying climate premiums for the NJ Department of Environmental Protection; STORM Act participation compressed yields while geospatial flood risk stayed unpriced.
  • Forecast 5-year justice-impacted returnee populations with ARIMA models against regional labor demand.

Projects

• osmo — His self-hosted personal AI agent. Its identity and memory are plain files in a git repo on hardware he owns.
  • Runs on an always-on Linux box and his Mac; an autosync job commits, merges, and redeploys between them every five minutes.
  • Has 24 tool servers: mail, calendar, coursework, GitHub, a browser, a code sandbox, a scheduler, health data, iMessage, and more.
  • A model router picks a model tier for each turn and scores 35/35 on its routing evals.
  • Its memory indexes 11,532 past AI conversations from five providers, with 0.974 recall@8 on its retrieval eval.
  • Reads freely, but sending as Aryan, spending money, or publishing passes a code-enforced approval gate.
  • Reachable by email or iMessage at [email protected]. Aryan approves its replies to anyone else.
• the tailnet — The private network of machines that Osmo and this site run on.
• markets on aryan — Prediction markets on Aryan’s own steps, sleep, and Osmo usage, settled against public telemetry.
  • Until shared trading exists, odds come from how often he has hit each line before. Aryan never trades.
• trade the knight — The exchange behind Rutgers’ first live algorithmic-trading competition, in April 2026.
  • A full-stack Rust exchange. Its matching core runs 5.5M orders per second with p99 latency of 0.85 microseconds (single thread, in-process, no network). The live competition ran at about 420 orders per second, limited by the bots and the API layer.
  • Built on Postgres, with Dockerized bot sandboxes and a Next.js front end; 90 participants competed.
  • Its predecessor was The Scarlet Exchange (2024): a simulated order-book exchange with participant bots and a Flask/React front end, built for club competitions.
  • The live competition had 90 participants and 54 Dockerized bots across spot, FX, bonds, options, and prediction markets.
• rl trading swarm — A closed simulated economy (stocks, options, bonds, a central bank and a treasury) with an evolving population of neural-network traders, used to test policy shocks and train RL agents.
  • On 8 seeds × 20 simulated years, 14–16 of 17 stylized-fact checks pass on every seed (fat tails, no return autocorrelation, long memory in volatility), with zero accounting invariant violations. The realistic dynamics come from designed fundamentalist and sentiment traders, not emergent behavior.
  • Paired policy experiments: a two-year +500bp rate peg drops the index about 30% (p = 0.002); a permanent −30% earnings shock drops it 21%; Tobin taxes of 10–50bp don’t reduce turnover.
  • A PPO agent beats every baseline on 48 held-out seeds. With its operating costs switched off it earns +7.1% a year at a Sharpe of 1.7; with normal costs every policy loses money.
• knightgym — A Gymnasium/PettingZoo reinforcement-learning environment built on the Trade the Knight matching engine, with a reward-hack audit.
  • In the audit, teammate wash trades under last-trade marking inflated marked equity by $1,077 per episode while true equity fell. Team-level self-trade prevention plus mid marking fixes it.
  • Built on the same Rust matching core: 5.5M orders per second, p99 0.85 microseconds.
• hospitalgym — A hospital operations twin built from public data: build the twin, calibrate a baseline, simulate, optimize. Every result is simulated; none is a measurement of the real hospital.
  • The first hospital is a model of Robert Wood Johnson University Hospital, New Brunswick, built from CMS, NHAMCS, and BLS data. Of its 129 hospital-specific attributes, 89 are measured, 6 come from literature, and 34 are assumed, each with a source tag and citation. It is not affiliated with the hospital.
  • The calibrated baseline reproduces the hospital’s published figures on 64 held-out seeds: median ED time for discharged patients 203 minutes against 206 published, and 0.91% left without being seen against 1%.
  • A batched discrete-time simulator on PyTorch tensors (5-minute ticks) runs one hospital for the 3D viewer or thousands in parallel for policy search.
  • In simulation on held-out seeds, a tuned rule-based plan cut mean door-to-provider wait by 17.0% (95% CI 9.1 to 26.2) and boarding hours by 59.6% at 4.0% lower labor cost, with every constraint met. It holds in 20 of 24 calibrations that public data cannot tell apart, and it misses a constraint under a flu surge.
  • The PPO agents have not beaten the tuned rule-based plans, and the report says so.
  • Built at RU Health Hack, October 2026. A recorded snapshot of the web app is at cantsoar.github.io/healthhacks26.
• poscere — A telehealth virtual-patient simulator for health-professions students: they interview an AI patient while a chart fills in, then get rubric feedback tied to quotes from the transcript. Built with one teammate at RU Health Hack, October 2026.
  • Code, not the model, decides which facts the patient releases, from each fact’s tier, the rapport the student has built, and the difficulty level. A second model checks every reply against the case and flags anything outside it.
  • 63 learner roles and 81 seed patients, all defined as data with no patient-specific code. A staged generator drafts a new patient in about 80 seconds. No seed patient has been reviewed by a clinician.
  • Care Journeys take one patient through several professions with a shared patient state, provenance on every entry, and SBAR handoffs that are scored.
  • Every stored visit is searchable from a page, a REST endpoint, and an MCP endpoint. It runs as one Next.js app with SQLite, Claude models on Amazon Bedrock, and ElevenLabs or Polly for the patient’s voice.
• aequus — Markets on any real-world quantity: priced by a virtual AMM, kept honest by bonded challenges and blind validator votes.
  • Solidity contracts (market, vAMM, vault, dispute pool, validator voting) deployed on the Base Sepolia testnet, with 95 contract tests after a security-fix round.
  • Grew out of an earlier tool that helps businesses hedge operational risk with Kalshi and Polymarket contracts.
• gto poker bot — A game-theory-optimal poker bot trained with fictitious play.
  • Maximizes expected value against modeled opponent ranges.
• electrostatics engine — A C++ physics engine that simulates charged particles and solves PDEs numerically.
• football greeks — A neural marked point process that simulates NFL games snap by snap to price team- and player-yardage contracts and their Greeks, tested against Kalshi.
  • Snap by snap, it uses a play-call net, a mixture-density network over outcome and yards, a pace net, and a rule engine.
  • Held-out log-likelihood 2.8476 against baselines at 2.8579 and 2.8850.
  • On 2024–25 games, its 50%, 80%, and 95% intervals cover about 48%, 78%, and 94%.
  • Paper-traded Kalshi team totals at +$0.115 per contract on one-minute candles over 49 games. On 16 games recorded live with real quotes it made +$0.05 per contract with a one-minute fill delay (95% CI −0.03 to +0.13): inconclusive.
  • Player-yardage models: tied with the market on QB passing; a second version cut the market’s lead on rushing and receiving by about 60%, but no trading rule made money.
  • A neural SDE fitted to the simulator gives autograd Greeks that agree with the Monte Carlo Greeks while more than 15 minutes remain.
• fantasy — Plays fantasy football on Sleeper as maygaxx.
• normal periods — Are irrational Kontsevich–Zagier periods normal numbers? Nobody knows even for π. This is a research direction with no result yet.
• thermal graph net — A PyTorch graph network that prunes cold edges and grows new nodes as it learns.
• true markets buildathon — Built live BTC/USDC trading at the True Markets buildathon in New York, June 2026.
  • A cross-exchange lead/lag arbitrage strategy and several market-maker versions.
  • Also simulated oracle-less markets that settle by Dutch auction, which later became Aequus.
• imc prosperity 3 — Team entry in IMC’s Prosperity 3 trading competition, 2025.
  • Pairs and hedging strategies with LSTM, XGBoost, and HMM models, plus brute-force and Nash-equilibrium solvers for the manual rounds.
• k_lib — A knowledge graph of quant research papers.
  • Ingests 52 sources (journals, arXiv q-fin, and bank research blogs) with section-level embeddings and a FastAPI and React interface.
• hyperbolic citations — Embeds the arXiv math and physics citation graph in hyperbolic space. The goal is to find papers that should be connected but aren’t. In progress.
  • A Rust parser feeding a PyTorch/geoopt Poincaré-ball model. The model hasn’t trained yet because the citation edges aren’t joined in yet.
• greeks from jacobians — Recovers option-style deltas and gammas for AAPL against sector ETFs and VIX from a neural network’s Jacobians, checked against a second-order Taylor fit.
• self-growing network — A network that detects a task change from a spike in its gradient norm and grows a new hidden block so it doesn’t forget earlier tasks.
• nets from scratch — Dense, recurrent, and LSTM networks in pure NumPy, with backpropagation derived by hand (2023).
• nostradamus — A 3D map of New Jersey that drills from state to county to all 2,175 census tracts to individual buildings, with live flights and satellite orbits.
• mati — An iOS app for drawing in augmented reality by tracking your hand.
• barclays datathon — First place at the BITS/Barclays Datathon (2025) with a buy/sell stock classifier.
  • Combined market data with FinBERT news embeddings, an SVM on technical indicators, and a retrieval-augmented transformer.
• ’95 miata — A 1995 NA Miata, bought to learn manual.
  • Bought it to learn manual, then rebuilt the engine.
  • Later sold it.
• ’07 bmw 328xi — A 2007 BMW 328xi (E90 sedan, N52 inline-six). The car he learned how cars work on.
• health telemetry — Steps, sleep, and heart rate stream from his wearable into public data and the market oracle.
• daily loop — A daily shortest-route puzzle. The server checks every route against the true shortest loop.

Interests

• coding — Python, Rust, C++, Java, Postgres, Docker, gRPC, AWS, and PyTorch.
  • Has been coding for over a decade.
  • Was posting coding tutorials on YouTube by 2017, starting with an introduction to HTML5.
  • Runs his own Linux servers on a private Tailscale network.
• trading — Trades just about everything, and builds trading systems.
  • Trades literally everything.
  • Organized “The Kalshi Games” for the Rutgers Quantitative Finance Club in December 2025.
• math — Stochastic calculus and algebra, mainly.
  • Likes stochastic calculus and algebra.
  • Coursework spans analysis, algebra, PDEs, econometrics, and the MSMF quantitative finance track.
  • Also interested in measure theory, option pricing, ring theory, and combinatorial game theory.
• football — Seahawks fan and fantasy player. Also a modeling project now.
  • A dedicated 12 (Seattle Seahawks fan).
• cars — Learned on a BMW, rebuilt a Miata engine.
  • Builds and races cars.
• optimization — Currently life-maxxing, with the data published.
  • Currently life-maxxing.
  • Publishes his own health telemetry so anyone can check it.
• training — Martial arts, lifting, and running.
  • Holds a third-degree black belt and trains MMA.
  • Lifts on push/pull/legs and Arnold splits, with jump rope for conditioning.
  • Once ran through severe Quebec cold around New Year’s.
• curiosities — Chess, cubing, languages, and tutoring.
  • Plays rapid chess and solves the Rubik’s cube with CFOP.
  • Reads German literature.
  • Has tutored STEM privately for more than 500 hours.

Elsewhere

• https://github.com/cantsoar
• https://www.linkedin.com/in/aryan-malik-xd/
• https://aryanmalik.web.app
• https://www.youtube.com/@aryanmalik2404