A person signs off.
Four-eyes review taught me that the check is the product. AI drafts and flags; people approve what matters.
About the founder
I spent two years checking the numbers at Goldman Sachs. Now I build AI that does the checking.
Arakyss builds AI into how trade and finance businesses run, with a person in charge of every decision that matters. That habit comes from the desk: at Goldman, the job was checks that had to be right, every day.
The path
The path so far.
-
Oct 2026 – now
Arakyss · founder
An AI studio that connects the systems a business already uses into one live picture and puts AI to work on it, with a person approving what matters: MESA. Every engagement starts with the 5-Day Build. I built this site too: a live 3D world where a simulated MESA runs an estate and asks before it acts.
-
Mar 2026 – now
Tazah AGI · founder, commodity intelligence
- Tazah AI: an assistant for a commodity trading desk. Ask about prices, buyers or a deal in plain language; it answers from live data, then drafts the contract, buyer email or client report for approval.
- A news-intelligence pipeline: AI reads 80 sources a day, pulls out the events that move prices, and keeps market views, each scored for confidence and paired with a test that would prove it wrong.
- ~PKR 70M a year found in a feed manufacturer's purchasing, by reconciling its inventory line by line against control totals and market prices.
-
Aug 2026 – now
Systematic research · personal project
- A systematic strategy for Pakistani equities (momentum picks with a trend switch into T-bills), run nightly and paper-traded, with signals sent to Axiom, an iPhone terminal I built in SwiftUI.
- Built to be falsified: the timing rule beat all 30 randomised alternatives while stock selection matched chance, so the edge is in timing, not picking. A look-ahead bias worth +0.32 Sharpe was found and removed.
- Modelled what frontier markets break: a maximum-likelihood volatility estimate for stocks that hit daily price limits (standard estimates missed up to 35% of the risk), and a market-impact cost model.
- Killed what did not hold up: machine-learning stock ranking lost to plain momentum in walk-forward tests (Sharpe 0.37 against 0.94).
-
Jan 2024 – Mar 2026
Goldman Sachs · Client Reporting Analyst, private markets
- Delivered same-day capital call and distribution notices to more than 10,000 investors in private-markets funds.
- Built a validation tool, "Investor Health Check", that cut processing cycle time by about 30%.
- As the Checker in four-eyes review, caught about $2.4M in errors by reconciling notices against investor-level data.
- Ran about six months of user-acceptance testing for the firmwide move to a new notice platform; go-live came early.
-
2023
Purdue University
BSc in Econometrics and Quantitative Economics, with a minor in Finance. Dean's Honor List.
How I work
Four habits from the desk.
Working before it's big.
One job, live on your own data, in five days. You see it work before anything larger is agreed.
Try to break it.
Placebo tests, walk-forward tests, reconciliations against control totals: if it can be proved wrong, I want to find out first.
Honest about the tools.
I build with AI coding tools. That is how five days is enough, and the time goes on your workflow.
What I work with
The tools behind the work.
AI
AI agents and tool use (MCP), Claude and OpenAI APIs, grounding and human-approval workflows, news-to-event extraction, on-device AI on Apple platforms.
Quant and statistics
Backtesting, walk-forward validation, placebo and bootstrap tests, event studies, maximum likelihood, volatility and market-impact models.
Data
Python, SQL, PostgreSQL, DuckDB, data-quality checks, PDF extraction and reconciliation against control totals.
Apps and web
TypeScript, React and Next.js, three.js and WebGL, Cloudflare; Swift and SwiftUI for iPhone, iPad and Mac.
Work with me
Start with one job, live in five days.
Tell me the job your team repeats every day. I'll come back with a plan and a fixed price.