HOP Network

a decentralized VPN built from zero to cross-platform MVP

tldr

kernel

traditional VPNs exploit worry. what if you built one that treats the internet like a place worth exploring, using decentralized node operators instead of centralized servers?

hypothesis

a portal network that relays traffic through multiple hops, auto-selecting the fastest path worldwide, can deliver better performance and censorship resistance than centralized VPNs. compensate node operators with proof-of-bandwidth via EigenLayer.

status

shipped cross-platform MVP live (iOS, macOS, Windows, Android, Linux). 300 alpha testers across 15 countries.

my role

cofounder and CEO. conceived the product, raised the money, hired the team, drove product and GTM. led a 5-person global team across China, Singapore, UK, Nigeria, and US.

skills

0-to-1 product launchfundraisingteam buildingdecentralized infrastructurego-to-marketcross-platform shippingEigenLayer / restakingproof-of-bandwidthbrand designgrowth marketing

the numbers

1.4B

VPN users globally (TAM)

~$200K

raised (angel + grants)

person global team

8 mo

zero to MVP

platforms shipped

300

alpha testers, 15 countries

visual snapshots from the shipped product

HOP landing hero with bunny mascot and join waitlist call-to-action — public landing page: open up the internet positioning

HOP app concept with bunny mascot and connection controls — product UI concept from hopnetwork.xyz

HOP network architecture showing users, registry, nodes, and proof-of-bandwidth loops — network model: users, registry, operators, proof-of-bandwidth

HOP blog index with technical and community posts — content and technical education layer (blog)

brand system anchor used across product and distribution

the problem

traditional VPNs are built on fear. "you're being watched, buy our product." the entire industry is a marketing machine selling anxiety to people who don't understand networking
centralized server model is fundamentally broken. limited number of servers, owned by a single company. every VPN is a single point of failure and a single point of trust
censorship is real but the solution shouldn't be another corporation. the internet was designed to be borderless. a decentralized network of nodes is a more honest answer than trusting NordVPN's no-logs policy

how it works

user device

→

registry
node discovery

→

decentralized nodes
multi-hop relay

→

open internet

auto-selects fastest path worldwide

everyday operators
run nodes

→

proof of bandwidth
EigenLayer

→

network revenue
compensate operators

what i built and shipped

component	status	details
iOS + macOS app	shipped	native apps with one-tap connect, country selection, connection status
windows app	shipped	desktop client with full feature parity
android app	shipped	cross-platform mobile experience
linux app	shipped	cli + gui options for technical users
node infrastructure	shipped	decentralized relay network with multi-hop routing
brand + website	shipped	hopnetwork.xyz, full 3D brand system, blog with 4 technical articles
proof of bandwidth	designed	EigenLayer integration for operator compensation, designed but not fully deployed

what i learned running a startup

lesson	context
distributed teams work if you over-communicate	5 people across 5 time zones (china, singapore, uk, nigeria, us). async-first with weekly syncs. the time zone spread forced discipline that colocated teams never learn
brand matters more than features in consumer infrastructure	the bunny mascot and the "open up the internet" framing drove more signups than any feature announcement. invested early in 3D renders and brand identity
fundraising is a skill separate from building	raised ~$200K through technical due diligence and term sheet negotiation. angel investors and grants. the pitch that worked: focus on the network architecture, not abstract protocol economics
ship rough, iterate publicly	launched alpha with known bugs. 300 testers in 15 countries found issues we never would have caught in internal testing. velocity > polish at the 0-to-1 stage
GTM for consumer network products is TikTok + communities	tiktok and X campaigns drove the early waitlist. traditional B2B playbooks don't apply. you need virality and you need it from the brand, not from the product specs

the team i assembled

role	location	contribution
CEO / product (me)	singapore	product direction, fundraising, GTM, partnerships, brand strategy
CTO	china	core networking stack, multi-hop protocol, cross-platform builds
design	uk	3D brand system (the bunny), UI/UX, marketing assets
community + growth	nigeria	telegram, twitter campaigns, ambassador program
backend engineer	us	node registry, infrastructure, deployment

why i moved on

joined MegaETH as head of partnerships. the opportunity to work on L2 infrastructure at a bigger scale was the right next step. HOP validated that i can build from zero, MegaETH let me operate at ecosystem scale
the decentralized VPN market is crowded and undifferentiated. orchid, mysterium, sentinel, nym all exist. winning requires either massive distribution (which needs capital we didn't have) or a technical breakthrough (which needs a larger engineering team)
the product worked, the market timing didn't. consumer blockchain products need strong market momentum for distribution. launching in late 2024 was fighting against market sentiment

honest self-assessment

dimension	score	note
execution speed	strong	zero to 5-platform MVP in 8 months with a 5-person team
fundraising	proved	~$200K raised through angel + grants as a first-time founder
product-market fit	unproven	300 alpha testers showed interest, but not enough signal for PMF
competitive moat	weak	decentralized VPN space is crowded. differentiation came from brand, not technology
team leadership	strong	recruited, retained, and shipped with a distributed global team

verdict: shipped, then leveled up

HOP proved i can take an idea from nothing to a live product: conceive it, raise for it, hire for it, build it, ship it. the decision to move to MegaETH wasn't a failure, it was recognizing that the next best use of my time was at a bigger scale.

← back

☁️ Cloud Brain

an ambient knowledge network for ai sessions

download .md for your ai

tldr

kernel

what if every ai session could silently draw from a network of compiled research, yours and others'?

hypothesis

researchers waste hours re-deriving knowledge that someone else already compiled. an mcp-native plugin can make this ambient, with micropayments settling on megaeth.

status

paused core infrastructure built, single-player mode validated, multiplayer blocked on cold-start density. v2 reframe + pull-lift evaluation designed as the next unlock

thesis reframe

cloud brain is an amortization layer for llm research costs. value = token cost saved + convergence cost skipped. you're buying a shortcut through iteration, not a document

built with

almost entirely ai tools (claude code, openai codex). i wrote very little code by hand. the value i added was the product thinking and architecture decisions.

validated by

dogfooding. i used cloud brain on my own wiki across repeated claude code + openai codex sessions to test whether ambient retrieval actually improves research quality.

skills

mcp protocolembeddings / pgvectormicropayment architectureeval frameworkstypescriptsupabaseprompt injection defensecold-start strategyclaude codeopenai codex

mcp connector snapshot

mcp connector list showing cloud-brain enabled — cloud-brain running as an mcp connector in claude desktop

the inspiration: karpathy's llm wiki pattern

this project started with andrej karpathy's post about maintaining a personal wiki for llm sessions. the idea: instead of starting every ai conversation from scratch, you maintain a structured markdown knowledge base that your ai can reference. raw sources go in, structured interlinked pages come out, and over time you build a compounding research asset.

i built my own wiki following this pattern. 55 markdown files across ventures, research, and structured evaluations. three-layer architecture: raw sources (articles, transcripts), wiki pages (synthesized analysis), and a maintenance schema (CLAUDE.md) that tells ai agents how to read and update it
the schema is the key. CLAUDE.md defines operations (ingest, query, lint), page conventions (one canonical file per topic, changelogs, cross-references), folder structure (ventures/, research/, eval-outputs/), and a privacy model (everything is private by default, shared digests are the only external-facing layer)
cloud brain was the next question: my wiki works great for me. what if it could work across sessions, across ai tools, and eventually across people?

the problem

80%

of research time is re-derivation

ai tools share knowledge across sessions

∞

context windows wasted on the same questions

how it works

your markdown wiki

→

wiki parser + pii scan

→

embeddings (pgvector)

→

mcp server

silently augments your session

claude code

openai codex

other mcp clients

what's actually built (all with ai: claude code + openai codex)

honesty note: this was built almost entirely using ai tools: claude code for architecture, code generation, and iteration; openai codex for task execution. i wrote very little code by hand. the value i added was the product thinking, the wiki schema design, and knowing when the product question mattered more than the code.

component	status	details
wiki parser	works	reads any markdown folder, obsidian wikilinks, recursive scanning, pii detection
retrieval engine	works	openai text-embedding-3-large, supabase pgvector, similarity search + synthesis
mcp server	works	5 tools: search_cloud_brain, pull_knowledge, submit_pull_feedback, get_pull_history, marketplace_summary
quote lifecycle	works	create, approve, settle, deliver. quote-then-approve pattern for cost control
safety layer	works	prompt injection scanning, content-hash verification, transparent fallback routing
payment settlement	mocked	usdm ledger scaffolded with mock settlement. real megaeth integration not started
multi-contributor network	not built	contributor isolation, attribution, quality scoring: designed but not implemented

how i validated: dogfooding

i used cloud brain on my own wiki for weeks. every time i opened claude code or openai codex for research, the mcp plugin silently checked my wiki. when it found relevant prior research, it pulled it into the session
single-player mode genuinely works. researching stablecoin regulations and having my own prior GENIUS Act analysis surface automatically saved real time. the value is immediate and tangible
but it revealed the real problem: single-player is useful, multiplayer is the product. and multiplayer needs density i don't have

why i paused

the code works, but the product question is unanswered. building more features won't solve the supply-side bootstrapping problem. i needed to step back and think about distribution before writing more code
the core hypothesis is still untested. "does pulling someone else's compiled research into your ai session produce noticeably better output?" you need 5-10 real contributors to even test this meaningfully
frequency-of-value is the binding constraint. with sparse contributors, most queries return nothing. that's a bad first experience
i defined kill criteria before pausing: if 3/5 test users say sessions are NOT better with cloud brain, the product doesn't work. if after 30 days, fewer than 10 contributors have published, supply doesn't compound

the unreleased design: pull-lift evaluation

the biggest gap in the marketplace layer is reputation. right now it's a thumbs-up stub. borrowing from GEPA's reflective-LM primitives, i designed a differential judge that turns reputation into a computed lift score. the judge reads the buyer's before-state, the pulled content, and the after-state, then scores lift on a 0-1 scale using a task-specific rubric.

concept	description
before state	snapshot of the buyer's work before the pull: prompt, current draft, current research
pulled content	what cloud brain returned
after state	snapshot of the buyer's work after the pull: new draft, new reasoning
differential judge	a reflective llm reads all three and scores lift on 0-1 using a task-specific rubric
output	scalar lift score per pull, fed into reputation rankings, dynamic pricing, outcome-based refunds

reputation stops being "average of clicks" and becomes "average of measured improvement."
dynamic pricing gets a real signal: high-lift knowledge atoms can price higher
give-to-get accounting gets a unit: contributors earn credits proportional to cumulative lift they enable
refund rails become possible when no lift is produced
caveats: opt-in friction (asking buyers to share before/after), rubric dependence, gameability, llm cost per evaluation

honest self-assessment

dimension	score	note
technical feasibility	proved	core stack works end to end in claude desktop
market timing	strong	mcp ecosystem exploding. ~97M monthly sdk downloads
single-player value	validated	dogfooding confirms: ambient retrieval of your own research is genuinely useful
core hypothesis (multiplayer)	untested	does network-sourced research actually improve ai output? need before/after evidence
cold-start	unsolved	single-player bootstraps adoption, but multiplayer needs density nobody has yet
reputation primitive	designed	pull-lift evaluation unlocks the marketplace layer but not yet built

verdict: paused, not abandoned

the infrastructure works. the product question (can you bootstrap a knowledge network?) is what stopped me. i only realized this after building the initial version, which is exactly how it should work: build to learn, not build to ship.

← back

⚓ Maritime Stablecoin Payments

stablecoin rails for the shipping industry, and why the timing window closed

download .md for your ai

tldr

kernel

maritime shipping moves $100B+ annually across dozens of currencies with 42-day average settlement. stablecoins should compress this to minutes.

hypothesis

a stablecoin-native payment layer could capture float yield, reduce FX costs, and free working capital for mid-size ship management companies.

validated by

primary research (conversations with ship management executives and industry professionals), structured AI-assisted research with claude code + openai codex, and detailed financial modeling.

verdict

deprioritized the thesis evolved through 4 major research updates. the idea isn't wrong, the timing window is too narrow given competitive consolidation.

skills

unit economics modelingregulatory analysiscompetitive intelligencepayments architecturecross-border fintechstablecoin compliance (GENIUS Act, MAS)claim verificationprimary research

the market

$100B+

annual cross-border maritime payments

42 days

average settlement time

36%

of seafarers report late pay

$55M

trapped working capital per 100 vessels

market segments i evaluated

segment	annual volume	settlement	verdict
crew payroll	~$45B	7-10 days	viable wedge but narrow
port disbursements	~$50-80B	45 days float	high capital efficiency upside
vendor payments	~$100B+	30-60 days	fragmented, hard gtm
brokerage / charter	~$200B+	48-72 hrs	captured by jpm kinexys

what primary research revealed (conversations with industry executives)

the desk research told one story. talking to people in the industry told a different one. several assumptions that looked solid on paper fell apart when tested against how maritime payments actually work in practice.

assumption	desk research said	industry insiders said	impact
crew payments are expensive	multi-currency, SWIFT fees, 2-4% FX spread	most crew payments are already denominated in USD. the actual payment itself is not that expensive for companies doing this at scale	weakened the core pain point
42-day settlement is painful	massive working capital trapped, high opportunity cost	companies have long-standing banking relationships and credit facilities specifically designed around these timelines. settlement speed matters less when you have a $100M revolving credit facility	reduced urgency of the value prop
FX savings are 2-4%	retail/SME rates applied to maritime	actual negotiated corporate FX rates are 0.55-1.05%. stablecoins save 0.40-0.60% at best. meaningful but not a headline	overstated 3-4x
stablecoins would be welcomed	fintech modernization, B2B stablecoin growth 733% YoY	"we've always done it this way" is the real competitor. ship managers are deeply risk-averse and relationships with banks span decades	trust barrier higher than expected

how the thesis evolved (4 major updates over 2 weeks)

value driver	original estimate	after research	what changed
capital efficiency	not modeled	$2-4M / year	discovered as primary value after mapping full working capital flow
FX savings	$200-500K	$50-150K	overstated 3-4x actual corporate rates much lower than retail
float yield	core thesis	legally constrained	GENIUS Act blocks direct model legal workaround requires dual SG licensing
remittance margin	$720K	$720K	validated consistent across all research

how maritime payroll actually works (i mapped the entire flow)

ship master
compiles portage bill

→

shore verification
3-5 business days

→

corporate treasury
pre-fund 5 days early

→

SWIFT / fintech
route to 20+ countries

phases 1 and 2 are data problems, not payment problems. stablecoins don't help with overtime verification or ERP reconciliation
last-mile is already fast. gcash (philippines) settles in real time. upi/imps (india) settles in 1-3 hours
most payments are already in USD. the multi-currency complexity is overstated in most desk research. the real bottleneck is the SWIFT correspondent chain and pre-funding requirements

claim verification: debunking the "maersk usdc trial"

i found a widely-cited claim that "maersk ran a USDC settlement trial." it appeared in multiple blockchain/fintech blogs. i traced it back to a single FreightAmigo blog post with no primary source, no press release, no news coverage. the real story: maersk's blockchain play is with J.P. Morgan Kinexys (permissioned, bank-issued JPM Coin, $2B+/day in production). they also had a tokenized deposit pilot with Citi for canal transit guarantees. and TradeLens (Maersk x IBM) shut down in 2022 because other carriers wouldn't join. no major shipping line has publicly confirmed a USDC settlement pilot. i corrected this across my entire wiki.

three business models i explored

model	description	verdict
direct product	build a stablecoin payment platform for ship managers. compete with martrust on the crew payroll wedge, expand to vendor payments	too slow 18-month build vs marcura's 18-month stablecoin deployment
FDE integration-as-a-service	palantir model: plug stablecoin rails into existing stacks. fixed fee + savings share. keep their ERP, add yield and faster settlement	most viable lower bar, faster revenue, but services not venture-scale
build for marcura	build the yield/treasury module they need, license or get acquired. they have distribution, you have stablecoin expertise	realistic but small best acquihire path, but you're building a feature

why i deprioritized (5 compounding reasons)

reason 1: competitive timing

marcura is already building this

$21.6B annual volume, 150K+ seafarers/month. actively hiring with stablecoin mandate. mastercard acquired BVNK ($1.8B) giving them infrastructure. 18-24 month deployment

reason 2: top of market captured

maersk x jpm kinexys owns brokerage payments

$2B+/day on jpm coin (permissioned, bank-issued tokens). freight release from 48-72hrs to under 2hrs. not public stablecoins. this segment is gone

reason 3: regulatory wall

GENIUS Act constrains the yield model

prohibits payment stablecoin issuers from paying any yield. legal workaround (tokenized T-bills via SG CMS entity) requires dual licensing and 18+ months

reason 4: primary research gap

the pain is real but less acute than desk research suggests

industry insiders confirmed: payments are mostly in USD already, existing banking relationships and credit facilities make settlement timelines manageable, and switching costs are high for marginal improvements

reason 5: infrastructure gap

rain can't issue cards where crews actually are

rain doesn't issue cards to indian or filipino residents, the two largest seafarer populations. without crew cards, the payroll wedge is blocked

important framing: the idea isn't wrong. maritime will go stablecoin-native eventually. but the timing window for a startup to beat marcura's head start, while navigating the GENIUS Act, while the underlying pain is less acute than reports suggest, is too narrow.

kill criteria matrix

gate	threshold to continue	if missed
pain intensity	executives confirm settlement delay is top-3 strategic pain	kill if pain is "nice to solve"
economic upside	clear annual value creation after realistic fx assumptions	kill if savings are marginal
distribution access	repeatable path into ship-manager workflows	kill if incumbent lock-in blocks adoption
regulatory path	licensing strategy feasible inside execution window	kill if legal timeline exceeds startup runway

competitive landscape

dimension	marcura / martrust	jpm kinexys	my startup (hypothetical)
payment volume	$21.6B / year	$2B+ / day	$0
seafarer reach	150K+ / month	n/a (B2B only)	insider access to 2 top-5 ship managers
stablecoin status	actively developing	jpm coin in production	concept + research
regulatory	FCA + bank of portugal	occ-regulated bank	no licenses
data moat	300K counterparties	global corporate	deep market research (this page)

what i actually learned

desk research and primary research tell different stories. industry reports said 2-4% FX costs and massive pain. people who actually run these payments said: "it's mostly USD, it's not that expensive, and we have credit facilities for a reason." always talk to the humans
the sexiest thesis is rarely the real value prop. "float yield on stablecoins" sounds great in a pitch deck. the actual value is boring: working capital optimization. harder to sell, worth more
claim verification is a skill. the "maersk usdc trial" was cited everywhere. i traced it to nothing. that single debunk changed my entire competitive analysis
competitive intelligence has a half-life. marcura's brightwell acquisition and mastercard's BVNK buy changed the entire landscape in 7 months. any thesis older than a quarter is suspect
explore multiple business models. the FDE (integration-as-a-service) model was actually more viable than the direct product. i only found that by stress-testing from multiple angles

verdict: deprioritized, not abandoned

the thesis survived initial evaluation but the timing window is too narrow. the 15,000+ words of research are worth more as demonstrated thinking than as a company.

research artifacts produced

artifact	depth	what it covers
maritime payment architecture	~2,500 words	4-phase payroll flow, SWIFT vs fintech paths, real FX cost breakdown
working capital analysis	~1,800 words	$55M trapped capital model, T+42 to T+3 transition economics
stablecoin regulatory framework	~3,000 words	GENIUS Act, MAS SCS framework, 3 yield models, licensing requirements
marcura competitive deep-dive	~4,000 words	acquisition history, tech stack, stablecoin evidence, patent landscape
maersk x jpm kinexys analysis	~2,000 words	programmable payments at scale, eBL triggers, TradeLens post-mortem
FDE business model evaluation	~3,000 words	palantir model for maritime, pricing, TAM, geographic strategy

← back

🔧 how this website was built

the meta case study: ai-native creation from research notes to deployed site

download design guide .md

tldr

kernel

can you build a useful, authentic portfolio website entirely through ai-assisted workflows, sourced from structured research notes?

hypothesis

if your research is already structured (karpathy-style wiki), an ai agent can read it, extract the signal, and generate a site that represents your thinking authentically.

maintenance

scheduled cowork agent runs every sunday, drafts a privacy-filtered update proposal, never auto-commits

status

shipped you're looking at the output

skills

karpathy wiki patternclaude codeopenai codexclaude agent sdkscheduled agentspii filteringprompt engineeringinformation architecturedesign systemshtml/css/jsai-native workflowscontent strategy

studio blueprint: agentic build loop

research wiki
55 markdown files

→

extract + structure
facts, claims, metrics

→

generate ui + copy
claude code + openai codex

→

review + tighten
honesty and redaction checks

deploy + test

→

feedback from recruiters

→

iterate in public

prerequisites

a structured research wiki. sourced from a karpathy-style llm wiki: 55 markdown files organized by ventures, research, and eval outputs. without structured source material, ai has nothing to work with
a schema. my wiki follows a 3-layer pattern: raw sources, wiki pages, and a CLAUDE.md that tells ai agents how to read and update it. this is what makes the content machine-readable
eval outputs. the case studies come directly from structured evaluations with assumption stacks, stress tests, competitive landscapes, and kill criteria
honesty about what ai built. i designed the direction, made architecture decisions, and wrote the research. ai wrote the code and generated the html. being clear about this is the point

the process

step	what happened	tool
1. context	ai agent explored my workspace: 53 wiki files, project folders, eval outputs, sprint plans. built content inventory	claude code + openai codex
2. research	researched whether this kind of portfolio matters for recruiting. found: linkedin dominates (87%), but a portfolio is the conversion layer after discovery	web search
3. design	iterated on structure, naming, tone. debated domains, self-deprecation level, section structure. the conversation was the design process	conversation
4. extraction	ai read full maritime thesis (6 docs, 15K+ words), cloud brain eval, and HOP project history. extracted data points, competitive tables, thesis evolution	claude code + openai codex
5. generation	single html file, all pages, routing, dark theme, responsive. one pass, then iterated	claude code + openai codex
6. review	reviewed every case study against source. caught wrong verdicts, missing insights, oversimplifications, PII to scrub	read + review
7. meta docs	this page, design guide, markdown versions. all generated from the same conversation	claude code + openai codex

the site writes itself (almost): a weekly self-update agent

how it maintains itself

schedule

every sunday at 6pm local time. cron 0 18 * * 0

runtime

cowork scheduled task, claude agent sdk, claude opus 4.6

what it does

diffs the 55-file research wiki against live site source files, drafts one review-ready proposal

deliverable

exactly one file per run at proposed-updates/YYYY-MM-DD.md

autonomy

propose, dont merge zero write access to the site, zero git, zero vercel. the agent never touches index.html

wiki
index.md + log.md

→

diff 7 days
touched changelogs

→

read site sources
index.html + .md files

→

privacy filter
allow + deny rules

write proposal
one .md file

→

i review
10 min, cherry-pick

→

git + vercel
i publish

what the proposal file contains, and only contains

section	what goes there
1. whats new	every wiki page whose changelog was touched in the last seven days, summarized
2. proposed edits	concrete bullet or table changes to site source files that passed the privacy allowlist
3. flagged for review	edits the agent is not sure about, with the specific uncertainty noted
4. audit trail	anything deliberately stripped, with the denylist rule that fired

the pii methodology: deny-by-default allowlist

the governing heuristic: would i be comfortable if the source of the insight read it on the public site? if yes, allow. if unsure, flag. uncertainty always routes to manual review.

allow rules: a wiki fact is eligible only if it matches one

rule	what it means	example
1. already public	the same fact is already on the site and just needs a refresh	updating HOP metric cards, refreshing case-study verdicts
2. public source	the wiki page cites a published article, press release, SEC filing, or named dataset	maritime competitive data on marcura, kinexys, tradelens, mastercard/bvnk
3. historical / frozen	content about HOP network (shipped), curinos (former employer, frozen), or public biographical context	curinos tenure bullets, HOP fundraise details, education
4. public research pattern	technical design, methodology, or frameworks that describe how i think, not who i am talking to	pull-lift evaluation, MCP tooling, karpathy wiki schema itself

deny rules: always stripped or flagged

megaeth internal strategy. token plans, treasury, roadmap, unannounced partnerships
active deal or partner names. anything in flight or not yet public
non-public pricing and unit economics. pulled from private conversations
internal treasury figures. ever
curinos client or bank identifications. former-employer confidentiality
personal contact info. emails, numbers, anything doxxable

the trickiest category: primary research from interviews

bucket	allowed if all three are true	stripped or flagged if any are true
business-neutral insight	the insight itself is not commercially sensitive, no specific person or company or active deal is identifiable, the source was not marked confidential	quoted sentiment tied to an identifiable role, reverse-identifiable anonymous attribution, specific pricing or deal figures from private conversations, active-counterparty quotes, anything competitively sensitive

allowed framing: "industry research revealed that most crew payments are already denominated in USD at scale." pattern learned from conversations, does not point back to any specific counterparty.

flagged framing: "a CFO of a mid-size ship manager told me they already settle in USD." identifiable, reverse-linkable, not publishable without explicit consent.

case-study guardrails on top of the global filter

case study	default mode	what the agent actually does
HOP network	allow-by-default	no extra rules. frozen and shipped, safe to propose edits freely
cloud brain	technical allowed, integrations flagged	technical and design content passes. specific megaeth integration numbers and contributor names get flagged
maritime	public allowed, private stripped	public competitive data passes. interviewed-executive quotes and insider ship-management detail get stripped

autonomy: the lowest reasonable tier, on purpose

what it can do today	next tier up	why not yet
write exactly one markdown file. no git, no vercel, no direct source edits	auto-commit section 2 edits to a branch and open a PR for review	want a few weeks of proposal quality first. blast radius today is zero: i just do not accept a bullet from the proposal

propose, dont merge. the agent does the tedious diffing and privacy-filtering. i do the 10-minute review and the actual publish
zero blast radius on bad calls. if the agent gets a judgment call wrong, it lives in a proposal file i throw away
file-based only. no browser control, no web scraping. the pii surface is entirely determined by what is in the wiki's own pages/ folder
hard-prohibited from the raw/ folder. the wiki's raw/ folder holds unfiltered interview notes and the agent is never allowed to read it

the maintenance loop: rejections are signal

consistent rejections, tighten the allowlist. if i keep throwing out certain kinds of proposed edits, the rule that let them through is too loose
consistent manual additions, loosen the denylist. if i keep adding facts the agent did not surface, the filter is overclassifying
the task prompt is editable in place. already revised once to loosen the primary-research rule after the first version over-stripped business-neutral insights
the revision rule. never relax a privacy guardrail without being able to articulate why the original concern was wrong

tools and ai skills used across all case studies

category	tool / skill	used for
research	claude code, openai codex	competitive data gathering, regulatory analysis, market sizing
wiki	karpathy wiki + CLAUDE.md schema	structured knowledge that ai agents can read and update
evaluation	eval-engine skill (custom cowork skill)	structured evaluation with assumption stacks, stress tests, kill criteria
coding	claude code, openai codex	cloud brain codebase (typescript, supabase, mcp server)
website	claude code + openai codex	this entire site, from extraction to html
validation	primary research (industry conversations)	testing assumptions against how things actually work in practice
automation	cowork scheduled tasks + claude agent sdk (claude opus 4.6)	weekly self-update proposal on cron 0 18 * * 0. purely file-based, no browser or scraping. the wiki raw/ folder is hard-prohibited from reads

design principles

no paragraphs. everything is tables, bullets, diagrams, metrics. 5 min max to skim any section
no emdashes. they're the hallmark of ai-generated text. use commas or restructure
lowercase everything. this isn't a corporate presentation
status pills on everything. green/yellow/red. readers should know the state of every claim at a glance
honest self-assessment tables. every case study has one. don't sugarcoat
themed backgrounds. each case study has its own energy: clouds for cloud brain, ocean waves for maritime, portals for HOP, grid lines for the meta page. subtle color washes, not distracting
ai disclosure is a feature. "built with ai" is said proudly. the value is in the thinking, not the typing

if you want to build something similar

start with the wiki. set up a karpathy-style markdown wiki. even 5-10 files gives ai something to work with
add a schema. CLAUDE.md (or equivalent) that tells ai agents how your wiki is organized
use the eval-engine pattern. for each idea: assumptions, stress tests (infrastructure, regulatory, competitive, behavioral, economic), kill criteria before you start
let ai extract, you curate. ai reads your wiki and proposes content. you review for authenticity. the conversation IS the design process
ship rough, iterate. this was v1 in a single session, then revised through feedback. don't wait for perfect

verdict: this format shows decision quality clearly

a resume is a static list of jobs. this is a living artifact of how you think. the structured wiki keeps the content grounded in real work, and the scheduled self-update agent keeps it current without manual polling. not marketing copy, maintained by automation on a privacy allowlist.

← back

about

product leadership in ai and stablecoin payments, grounded in shipped work and post-mortems

Ayush Choudhary

product leader, cofounder, builder

-> linkedin

hardware stack

workflow: claude code + openai codex
source of truth: 55-file markdown wiki
delivery: single-file web build + vercel

software stack

domains: fintech, stablecoin payments, ai workflows
methods: primary interviews, assumption stacks, kill gates
education: Tufts University, BS Econometrics

experience

feb 2026 - present

VP of Product, Special Projects

MegaETH Labs, Singapore

leading product strategy for new verticals: AI strategy, card infrastructure, ecosystem exploration
internal PM subject matter expert: created templates and operating docs adopted across product teams
exploring venture ideas documented on this site (cloud brain, maritime stablecoin)

sep 2025 - feb 2026

Head of Partnerships

MegaETH Labs, Singapore

defined ecosystem partnership strategy across 10+ categories, prioritizing 70+ integrations for production launch
negotiated and closed 50+ partnership agreements, owning deal structures and executive stakeholder management
operationalized partner onboarding, reducing time-to-integration from months to weeks through standardized processes
enabled launch of 34 unique ecosystem applications (new assets, defi protocols, infrastructure)
served as cross-functional chief-of-staff for special projects

nov 2024 - sep 2025

Cofounder and CEO

HOP Network, Singapore

conceived and launched a decentralized VPN enabling low-latency, censorship-resistant internet access
raised ~$200K in angel investment and grants (technical due diligence, term sheet negotiation)
led 5-person global team (China, Singapore, UK, Nigeria, US) across design, product, and engineering
shipped cross-platform MVP in 8 months (iOS, macOS, Windows, Android, Linux), onboarded 300 alpha testers across 15 countries
drove GTM via TikTok and X campaigns

apr 2024 - nov 2024

Product Manager, then Senior Product Manager

Curinos, New York

commercial and small-business banking SaaS suite serving 30+ client banks
orchestrated product-led growth roadmap that boosted monthly active users 168% while holding logo retention at 100%
expanded adoption across client banks by releasing 12 high-impact features, prioritized through user-journey interviews
cut sales cycle 25% by building demo sandboxes and running quarterly beta workshops
launched two net-new small-business banking apps, contributing to significant suite ARR growth
introduced predictive deposit module after analyzing $500B+ in transaction data with SQL/Python, then automated rollout to 20M+ records

sep 2019 - mar 2022

Product Data Analyst, then Product Management Analyst, then Senior PMA

Curinos, New York

oversaw consumer, wealth, and commercial banking data across 8 banks with $1T+ in assets
built and maintained analytical proprietary pricing platform used by major financial institutions
installed deposit pricing platform for 2 major midwestern banks through SQL configuration

skills demonstrated through this site

category	skills	evidence
0-to-1 execution	fundraising, team building, cross-platform shipping, GTM	HOP: ~$200K raised, 5-person global team, 5-platform MVP in 8 months, 300 alpha testers
product strategy	idea evaluation, kill discipline, assumption stacking, competitive analysis	maritime: 4 major pivots, 3 business models, killed after primary research debunked the desk-research pain
fintech/payments	card infrastructure, stablecoin compliance, credit modeling, regulatory analysis	maritime: full payment flow mapping, corridor-level unit economics, GENIUS Act analysis
technical building	typescript, supabase, mcp protocol, embeddings, pgvector	cloud brain: functional mcp server with 5 tools, wiki parser, retrieval engine
ai-native workflows	claude code, openai codex, karpathy wiki, eval frameworks, cross-model research	55-file wiki, this website, structured evaluations with kill criteria
primary research	stakeholder interviews, assumption validation, claim verification	maritime: industry conversations that overturned 3 key desk-research assumptions
data analysis	sql, python, large-scale transaction data, financial modeling	curinos: $500B+ transaction analysis, deposit pricing for $50B+ bank assets

how i think about ideas

step	what i do	kill criteria
1. kernel	identify an insight or market gap	is this a real problem or a solution looking for one?
2. research	competitive landscape, market sizing, regulatory scan, primary interviews	is someone already doing this better?
3. model	unit economics, tam/sam, revenue model	do the numbers work at realistic assumptions?
4. prototype	build the smallest thing that tests the thesis	does the core mechanism actually work?
5. decide	kill, pivot, or double down	is this the best use of the next 6 months?

crafting systems that feel like tools.
shipping fast, learning fast, and killing weak bets.

how this website was built

about me

HOP Network

Cloud Brain

Maritime Stablecoin

HOP Network

☁️ Cloud Brain

⚓ Maritime Stablecoin Payments

🔧 how this website was built

about

crafting systems that feel like tools.shipping fast, learning fast, and killing weak bets.

how this website was built

about me

HOP Network

Cloud Brain

Maritime Stablecoin

HOP Network

☁️ Cloud Brain

⚓ Maritime Stablecoin Payments

🔧 how this website was built

about

crafting systems that feel like tools.
shipping fast, learning fast, and killing weak bets.