• 0 → 1 products
  • Payments
  • Cross-chain
  • Agents in production
  • Mobile + web
  • New York
Founding Engineer · New York, NY

Marlon

I start from ambiguous projects and own the whole system.

Open to founding and senior roles in AI-native products or fintech. New York or remote.

github.com/marlon-ma
THE ONE-PAGER

The résumé

Now
Founding Engineer at Stack (formerly Hawala), since Jul 2025 · $3M pre-seed
Before
Northeastern AI Lab · NetEase · founded a photography SaaS
School
M.S. NYU, 2024–26 · B.S. CS Northeastern, top 1%
Base
New York, NY · open to remote
Download the PDF
Marlon Ma, one-page résumé
  • React Native
  • Expo
  • React
  • Custodial cards
  • On/off-ramps
  • KYC
  • Solana
  • EVM
  • USDC
  • Deno edge
  • Node
  • PostgreSQL
  • LLM orchestration
  • Tool use
  • RAG
  • Zod
  • ESLint
  • Migration-first
Founding Engineer · Jul 2025 – Present

Stack

Stablecoin neobank for emerging markets. Formerly Hawala.

Blank repo to live app.

MO

Blank repo → live consumer app

Commits owned at Stack

%

Of entire repo

Teams running on my AI platform

Pre-seed, joined idea-stage

s

Median AI support resolution

WHAT'S NEXT

AI-native fintech is not a chatbot bolted onto a form.

The product itself is the agent. Say “send tuition to my daughter at NYU” and it works out the route, the compliance, the rate and the timing. Ops scale with intelligence, not headcount.

01

Ameen

The first agent in the layer.

One agent doing what hundreds of people would. Nine languages. Every ticket becomes training data.

  • Support
  • KYC
  • Declines
  • Transfer issues
02

Ask Stack

Same engine, turned to the user.

Personal finance in plain English.

  • How much did I spend on coffee.
  • Send mom a thousand dollars for Eid.
  • Should I move savings into stablecoins.
03

Agentic Economy

The wallet for the next trillion users, who won't be human.

Agents pay each other. We build this for emerging markets while the West builds it for the West.

  • Stablecoins
  • Micropayments
  • Programmable limits
INTERNAL AI OS

AI platform

Not a chatbot.

Six teams use it every day. It answers in Slack, works a queue overnight, and stops before it can break anything.

The 3D office — a floor plan with agents standing on it, running against a demo gateway
THE OFFICE

You can walk through it.

Every agent has a desk. Open the room and see who is working, who is idle, who errored.

Renderer: Claw3D by LukeTheDev · MIT · vendored · demo gateway, fictional agents.

One agent · three jobs
#incidentsLive
Marlon8:04 AM
500s on /transfer since 07:40. Someone take it.
Debug agentApp8:04 AM
On it. Pulling Sentry, Supabase and Git in parallel.
Debug agentApp8:11 AM
Root cause found. I've drafted the PR. I don't write to production, so it needs one of you.
Root-cause PRDrafted
SentrySupabaseGit
Hours → <10 minTime-to-RCA
#opsLive
Sam Bauer4:12 PM
@agent delete the staging env and the old repo to cut cost.
Ops agentApp4:12 PM
Killing staging now. The repo has a commit from today and 3 open PRs, so I'd hold and archive it instead?
acme-platform · last commit 2h ago · 3 open PRs
Sam Bauer4:13 PM
good catch. archive it.
Zero auto writesSafety by construction
#supportLive
Amina11:58 PM
Where's my transfer? Sent it two hours ago.
Support agentApp11:58 PM
Pulled your ledger from the partner rail. It's clearing now, funds land within the hour.
Same agent, six languages
ArabicHindiMandarinSpanishFrenchVietnamese
9–38s medianTier-1 support, six languages
WHY IT IS BUILT THIS WAY

Whose computer is your company thinking on?

Every AI tool a company adopts moves its documents onto somebody else’s servers.

So this one gives every person a machine of their own, in the cloud, that never turns off.

  • One machine each

    Own container, own storage, own network. Nobody reaches anyone else’s.

  • Your key, your traffic

    Bring your own model key. The gateway logs that a call happened, never what was in it.

  • It works while you sleep

    Scheduled jobs and a board of cards, worked with nobody watching.

What it may touch
53admitted, across six toolkits
1,000+reachable through the integration platform

Deny by default. Every tool is reviewed in a pull request and pinned to the reviewed schema. Reads run; anything that writes stops for a human.

  • GitHub
  • Gmail
  • Google Calendar
  • Notion
  • Linear
  • Slack
Saad

UI-preview build · fixture data

Saad: Agents
AgentsYours, your team’s and the company’s in one roster, with what each may touch.
Saad: Memory
MemoryWhat the agent has been taught — readable and removable by whoever owns it.
Saad: Approvals
ApprovalsEvery write stops here first. One use, one action, time-limited.
Saad: Support console
Support consoleLive tickets grouped by theme, each with a diagnosis behind it. Replies stay drafts.
Saad: Connectors
ConnectorsWhich model thinks, where you reach it, where it may act. Keys stay in the broker.
Saad: Learning
LearningEmployees propose, the company reviews, then it ships. Nothing spreads on its own.
Saad: Rooms
RoomsShared spaces where a whole team talks to the company agent at once.
Saad: Tasks
TasksThe board agents actually work. A card is a prompt; a run has to end on it.
BUILT ON MY OWN · 2026

MOXIE

A personal finance agent that has to prove every number it says.

Reads your bank through Plaid, plans the month, answers “can I afford this”. Every number in an answer opens the math behind it.

Open in a real tab ↗
THE PART THAT MATTERS

An answer you can audit.

Press any amount and the calculation opens: the period, the coverage, the freshness, every transaction behind it. A number the tools cannot back is withheld.

  • One uncited figure kills the whole answer
  • The tools do the arithmetic; the model only narrates
  • Every figure is checked to the cent before it is shown
ExpoReact NativeNext.jsTypeScriptSupabasePlaidClaudemoxuse.com ↗
Moxie, the calculation behind a cited amount
CONCEPT · NOT SHIPPED

The agent plans. You press pay.

MoxieConcept

3 days in Lisbon in October. Keep it under $1,200.

Tue–Fri is the cheap window. This fits your flexible room and leaves $180.

Lisbon · Oct 14–17

TAP · JFK → LIS
$486
Hotel, 3 nights
$402
Food and transit
$132
Total$1,020

Pay with the Sapphire — 3× on travel, $30 back

  1. Checked October's flexible room: $1,200
  2. Compared 14 fares, held the cheapest
  3. Picked the card with the best travel rate
Moxie never pays. You do.
Plan a tripIt knows what is actually free this month before it books anything.
MoxieConcept

Restock for the week. The usual.

Fourteen things you buy most weeks. Whole Foods on Houston is cheapest today.

Basket · 14 items

Produce, 6 items
$28.40
Dairy and eggs
$19.10
Coffee, the one you rebuy
$16.00
Instacart total$104.60

Delivery 6–7pm · $12 under your weekly average

  1. Read six weeks of grocery receipts
  2. Kept the 14 items you rebuy every week
  3. Priced the basket at three stores nearby
Moxie never pays. You do.
Plan groceriesBuilt from six weeks of your own receipts, not a template list.
MoxieConcept

Booking the flight now. Which card?

Sapphire. Travel is 3× on it and your Amex is 1×, and the statement closes Friday.

On $486 of travel

Sapphire · 3×
+$29 back
Amex Blue · 1×
+$10 back
Debit
nothing
Difference$19

Statement closes Fri — this lands on the next cycle

  1. Found the three cards you hold
  2. Looked up each one's travel multiplier
  3. Checked which statement closes first
Moxie never pays. You do.
Pick the cardIt knows which cards you hold, so it knows which one to reach for.
Arc

Five years, one line.

  1. 2020

    Started at Northeastern. AI lab, first model training before the hype cycle.

  2. 2022

    Only undergrad in Northeastern's AI lab. Shipped YOLOv5-G to a factory QC line.

  3. 2023

    NetEase. Stable Diffusion in production for 800+ DAU.

  4. 2024

    Top 1% of cohort. B.S. CS at Northeastern. M.S. at NYU begins.

  5. Jul 2025

    First engineering hire at Stack (then Hawala). Blank repo.

  6. Now

    Live app, $3M pre-seed, AI platform 6 teams use daily.

BEFORE STACK

Everything before this

The only undergrad in the lab · Northeastern AI Lab · Sep 2021 – Jun 2024

Industrial defect detection

Designed YOLOv5-G for sub-2mm defects on a glass maker's QC line: +8.3% small-object mAP, precision 60% → 82%, manual inspection down 60%.

YOLOv5-GRepGFPN60% → 82% precision

Three years in a PhD lab · Northeastern AI Lab · Sep 2021 – Jun 2024

Medical image segmentation

Reproduced U-Net on hospital CT and matched the paper's Dice on a held-out set. Wrote the DICOM pipeline the lab kept using.

U-NetDICOMHospital CT

Software Engineer Intern · Beijing · Dec 2022 – Feb 2023

NetEase

Style-transfer service in PyTorch + Flask: inference 3–5s → under 1s at 800+ DAU. Put Stable Diffusion in the design team's workflow; feedback cycles 3 days → 8 hours.

PyTorchStable Diffusion−85% latency
Tools I reach for

What I build with

Mobile from blank repo

React NativeExpoReactNext.jsTypeScriptEAS release

Every consumer flow

Custodial cardsOn/off-rampsKYCReferrals

Cross-chain settlement

SolanaEVMUSDCCross-chain bridgingPrivy

Backend + data

Deno edgeNodePostgreSQLSupabaseRedisKafka

Two flagship AI agents

LLM orchestrationTool useRAGEval harnessesAgent safety + HITLOpenAI / Anthropic APIsPyTorch

Engineering bar as team grew

ZodESLintMigration-firstAWSDockerGitHub ActionsSentryPostHog

Let's build the next one.

Happy to walk through any of it. Email is fastest.

github.com/marlon-ma
Résumé (PDF)

P.S. Click anywhere.

New York, NY · Founding Engineer

© 2026 Marlon