← All sessionsHomeSearch
AI Sprints (Live Weekend Programs)·Building & Shipping Reliable AI·3:25:37

AI Sprint: Vibe Coding to Production — Day 1 (Deployment, DNS, and the Five Locks)

Sugam Guest mentor — head of engineering at Outskill ('we are serving right now hundred million users per month'); builder/engineer, 3x founder with an acquisition; 15 years in; his first time teaching a sprint session ('this is my 1st session') · Sumedha Sprint host ('I'll be hosting your sprints going forward') — opening, Q&A triage, and the community showcase segment
▶local recording  ·  LMS lesson

Session map

THE MINDSETTHE MECHANICSTHE PROOFThink for scale, ship for tenthe nobody-uses-it-yet correctionNo best stackcomfort, agent-compatibility, and the h…Git the system, GitHub the sh…The five locks + two prompt l…security for vibe-coded appsCommit → build → release → wa…the deployment loopShip cheap, then showthe <$10 receipt and the cohort showcase
The mindsetThe mechanicsThe proof
click a node — its card pops up (drag it anywhere, × to close)
Concept

The map reads left to right — the mindset flow into the mechanics, then into the proof. Click any node to open that idea here; every timestamp jumps into the recording.

The short version

  1. The sprint's day-1 thesis: 'building a prototype is completely a different thing than making something work on your laptop' — the gap between built-it and shipped-it is not the code, it's deployment, and the session walks the whole gap live: GitHub push → Vercel front → Render back → Atlas DB → Stripe → custom subdomain with auto-SSL.
  2. The mindset correction opens it: 'you don't have a scale problem. You have a nobody-uses-it-yet problem' — think for scale in the architecture prompt (one added line naming the deployment stack), then 'pick any boring managed option. Ship it.'
  3. Stack doctrine: there is no best tech stack — comfort + agent-compatibility + hiring pool decide; his 2017 Vue.js scar ('1 in 10 developers... I had to migrate the entire system to React') is the cautionary tale; default answer Next.js + Node.js, loosely coupled so scaling is a one-click instance upgrade.
  4. Security is five locks plus two prompt lines: no secrets in the repo (.gitignore), no keys in the frontend (AI browsers will crawl and find them), no exposed DB credentials/weak passwords, unguessable admin routes, rate limits — and every build prompt gets 'write a unit test case for each query' + 'no secret parameter exposed to public routes.'
  5. The deployment loop to memorize: commit → build → release → watch → rollback — continuous deployment defined by contrast with the SSH-and-YAML past; Vercel's instant rollback demoed; databases only roll back if you keep dev/staging/prod environments.
  6. Receipts throughout: his getout.ai runs a lakh (100k) users/day for UNDER $10/month on the boring stack — and the closing 45-minute showcase parades six cohort builds, including an AEO/GEO audit platform (PolarMirror) directly relevant to this KB's SEO priority.

The concepts

01

Think for scale, ship for ten: the nobody-uses-it-yet correction

He polls the room: how many real users — 'not you and not your friend' — does your last project have? The chat fills with zeros. That's the diagnosis.

Two opposite problems get conflated: a SCALE problem (traffic breaking your system) and a NOBODY-USES-IT-YET problem (a GTM/market question). 'Those are opposite problems, and they call for opposite decisions.' Most builders have the second while architecting for the first — 'building for a million while serving 10 is not ambition,' it's misallocated effort.

The resolution is temporal: think for scale at DESIGN time, cheaply — one added line in the vibe-coding prompt fixing the deployment stack ('deploy on Vercel, deploy on Render, deploy on MongoDB Atlas') and a real-user number held in mind — then ship the boring managed option immediately. The restaurant analogy sizes it: 10 guests at one table (anything works — Vercel, Render), 10,000 guests nightly (managed hosting, a real database, basic monitoring), a million (serious attention — and a problem you'll be glad to have). The counter-example is the pay-for-leaderboard site he shows: launched in 5 days, $17k collected, a million visitors — 'accidentally built' with zero scale planning, and his verdict: 'one month down the line, this platform will be down.'

Worked example · from the session

The intern receipt: getout.ai, Outskill's internal short-link platform, 'designed by 1 of my intern' with his scaling guidance — live stats shown on screen, lakh-a-day traffic, the proof that think-for-scale is a prompt-and-architecture decision, not a budget.

Why it matters

This is the session's frame: every later topic (stack, security, deployment) is executed at ship-for-ten cost while remaining scale-compatible.

People get this wrong

A serious product needs serious infrastructure from day one.

It needs scale-COMPATIBLE choices from day one (boring managed platforms, loose coupling) — actual scaling is a later, one-click-or-paid problem you earn.

Building a prototype is completely a different thing than making something work on your laptop.
You don't have a scale problem. You have a nobody uses it yet problem. Those are opposite problems, and they call for opposite decisions.
For your projects

The 'nobody-uses-it-yet' framing is a diagnostic worth stealing for client conversations — most 'performance' asks are actually adoption asks.

This idea elsewherebuilds onthe same 80/20 — product truth before machinery
Go deeper

In one line: Separate scale problems (traffic) from adoption problems (GTM); architect scale-compatible via one prompt line naming the deployment stack + a real-user number; ship the boring managed option now. 10 users → anything; 10k → managed hosting + real DB + monitoring; 1M → serious engineering, later.

'You don't have a scale problem. You have a nobody uses it yet problem... opposite problems, opposite decisions' ()

The one-line prompt addition pinning stack + deployment target — 'be specific about the deployment as well' ()

'Pick any of the boring managed option. Ship it. Don't think too much' ()

The $17k leaderboard site as anti-example: virality without architecture — 'this platform will be down' ()

'The gap is not the code' — twice; the agent ('the most advanced intern') wrote fine code; the gap is everything after ()

▶ Watch this taught:

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

What's the difference between thinking for scale and building for scale?

Thinking for scale is free: name the stack and deployment in the prompt, keep the real-user number in mind, keep front/back loosely coupled. Building for scale (Kubernetes, enterprise stacks) waits until real traffic — or real money — demands it.

02

No best stack: comfort, agent-compatibility, and the hiring pool

His costliest mistake in 15 years wasn't an outage — it was a language choice: Vue.js in 2017, '1 in 10 developers,' and a full paid migration back to React.

Three selection criteria replace the best-stack question: (1) YOUR comfort — 'whatever you are comfortable is the best tech stack for you'; (2) agent-compatibility — 'don't jump around to multiple tech stacks. Otherwise your vibe coding agent will confuse as well'; if no strong preference, 'go ahead blindly with Next.js and Node — it will work best with your coding agent'; (3) the HIRING POOL — choose languages where developers are findable, the Vue lesson.

His defaults with reasons: Next.js frontend (deploys to Vercel with CI/CD, logs, monitoring built in), Node.js backend on Render — SEPARATE repos and services, 'loosely coupled... if I want to scale it, I can scale the system very easily.' Databases by shape: document (MongoDB/Atlas), file (SQLite/Turso), relational (Postgres/MySQL) — plus Redis as the speed layer that carries either to 100M-request territory. Golang for genuine high-traffic infra (getout.ai). Enterprise stacks (Spring Boot, ASP.NET) 'when you have money in your pocket to afford a developer.' Mobile: React Native or Flutter for reach; Kotlin/Swift only when the app touches hardware (camera, GPU, face recognition).

Worked example · from the session

The famous-users tour anchoring each stack: Reddit on Django, Stack Overflow on Rails, early Facebook on PHP, WordPress on PHP — 'there is no best,' there are teams that shipped.

Why it matters

Stack anxiety is the most common beginner question in the room ('which tech stack should I use?') — the three criteria dissolve it into decisions a non-engineer can make.

People get this wrong

Scalability lives in exotic technology choices.

It lives in boring, well-staffed, agent-friendly choices arranged loosely — the exotic stack is how you end up unable to hire or migrate.

There is no best tech stack... whatever you are comfortable is the best tech stack for you.
For your projects

His agent-confusion point is real in your pipelines too — one stack per project keeps builder agents deterministic.

Go deeper

In one line: Stack selection = comfort × agent-compatibility × hiring pool. Default: Next.js + Node.js in separate services (Vercel + Render), DB by data shape (Atlas/Postgres/SQLite) + Redis for speed; Golang for high-traffic infra; enterprise Java/.NET only with budget; React Native/Flutter for mobile, native (Kotlin/Swift) for hardware access.

'There is no best tech stack... whatever you are comfortable is the best tech stack for you' ()

The Vue scar: 'really hard to find a developer... I had to migrate from Vue.js to React again, the entire system' ()

Loose coupling = one-click scaling: 'I just need to pay more in order to get more room' on Render ()

Redis's promotion: 'Postgres plus Redis... can handle hundred million users' traffic very easily' ()

Stick to ONE stack per project — agent confusion is a real cost ()

Vercel favored partly because 'Next.js is basically designed via our team only' — as heard; Vercel is Next.js's maker ()

▶ Watch this taught:

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

Why do front end and back end get separate repos and hosts in his default?

Loose coupling: each scales, deploys, and rolls back independently — 'if you have built everything in 1 repo, it will be very hard for you to decouple it again.'

03

Git the system, GitHub the shelf — and .gitignore the lock

'Why do I need to install Git if I already have GitHub?' — the question half the room was afraid to ask, answered with a shelf.

Git is the local SYSTEM: 'a written record for every change that you made for your recipe' — versioning that lives on your machine, free forever. GitHub/GitLab/Bitbucket are SHELVES: cloud hosts where the record gets pushed (Microsoft/independent/Atlassian respectively; hosting free, extras — code review agents, CI/CD pipelines — monetized). The four-command loop demoed live: git init, git add, git commit -m, git push (with the live hiccup — a failed push because the remote wasn't added — left in and fixed).

The security half is .gitignore: 'this file is very, very important' — it keeps env files, secrets, node_modules, and build artifacts OFF the shelf, so a repo accidentally flipped public (by you or your agent) doesn't leak keys. His move: ask the agent directly — 'add git ignore file in both the folders' — and verify env/example files stay local.

Worked example · from the session

The two-repo push: demo frontend and demo API repos created on camera, .gitignore protecting both env files, then Vercel and Render each importing straight from GitHub — the shelf as the deployment source.

Why it matters

Version control is the hinge of the whole deployment loop: no push, no build; no .gitignore, no secrets left to protect.

People get this wrong

GitHub is Git.

GitHub is one shelf Git can push to — the versioning system is local, free, and works with any shelf or none.

Git is a process... GitHub is a shelf where you host your code.
For your projects

Matches your lovable-project-setup GitHub-protection standard — this is the cohort-level version of the same rule.

Go deeper

In one line: Git = local version-control system (free); GitHub/GitLab/Bitbucket = hosting shelves monetizing extras. Working loop: init → add → commit → push. .gitignore keeps env files/secrets/node_modules off the shelf — agent-addable ('add git ignore file in both folders'), verified before the first push.

'Git is a process... GitHub is a shelf where you host your code' ()

All three shelves host free; 'these platforms are making money from those extra services' ()

The .gitignore protects against your OWN agent: 'if I make my code public by mistake or my agent have done that' ()

Agent-driven pushing works too: 'I can ask directly agent to do the end to end thing' — the terminal is his preference, not a requirement ()

▶ Watch this taught:

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

What belongs in .gitignore before the very first push?

Every env file (and env.local variants), node_modules, build outputs, logs — anything holding secrets or bulk that the shelf doesn't need.

04

The five locks + two prompt lines: security for vibe-coded apps

His attacker isn't a hoodie in a basement — it's an AI browser: 'it will go to each and every route... and get the keys for me.'

Five locks: (1) never commit secrets to GitHub (.gitignore); (2) never expose keys in the frontend — modern AI browsers can be told to crawl a site and harvest OpenAI/OpenRouter/Claude keys from rendered pages; (3) never expose DB credentials, and use hashed/strong passwords; (4) hide the admin — attackers probe /admin and /admin/login by default, so his super-admin route is unguessable; (5) rate-limit exposed endpoints, or a single chat endpoint becomes a DDoS bill.

Two lines get added to every build prompt: 'write a unit test case for each query' (which also covers SQL injection — 'you can mention SQL injection as well') and 'no secret parameter should be exposed to public routes.' Above the locks sits the agent-as-auditor move: 'audit the entire code base. What are the top risks that you see?' — his own live audit caught a refund-handling flaw mid-session and fixed it. And the homework, assigned for TONIGHT: open dev tools on your live site, Cmd+F for 'secret'/'key' — 'if you can find it, anyone else can find it'; optionally point an AI browser at your own site with the same harvest prompt attackers would use.

Worked example · from the session

The live secret-hunt demo on his own app: inspect → elements → Cmd+F across env-style names (Stripe secret key, webhook secret) — showing exactly what an attacker's copy-paste session looks like.

Why it matters

Vibe-coded apps ship with agent-written code nobody reviewed — the locks + prompt lines + self-audit are the review a non-engineer can actually run.

People get this wrong

Security means adding a security product.

For vibe-coded apps it means five configuration habits, two prompt lines, and an agent audit — all free, all before any product purchase.

The five locks (+ two prompt lines) Secrets .env, never in code Auth who are you Authorization what may you touch Validation never trust input Rate limits stop the flood Two prompt lines “apply the five locks; show me where each lives” The security floor for vibe-coded apps — cheap to ask for, expensive to skip
Five locks plus two prompt lines: the security floor for vibe-coded apps
If you can find it, anyone else can find it.
For your projects

The dev-tools key-hunt belongs in your site-health-audit checklist verbatim — it's a free check with a binary result.

This idea elsewherebuilds onsecurity delivered agentically — the checker-loop idea applied to deployment
Go deeper

In one line: Locks: secrets out of repos; keys out of frontends; DB credentials hidden + strong passwords; unguessable admin routes; rate limits on endpoints. Prompt lines: unit tests per query (incl. SQL injection) + no secrets on public routes. Agent audit: 'audit the entire code base — top risks.' Homework: dev-tools key-hunt on your own live site.

AI browsers as attack scouts: give ChatGPT Atlas/Comet-class browsers a harvest prompt and they'll walk every route ()

'Most of the hackers will come and see /admin... keep the URL so that no one can guess' ()

Rate limiting framed as cost defense: an unlimited endpoint is a DDoS invitation ()

The live self-audit caught a real bug: 'a partial refund... reported revenue' mismatch, fixed in the webhook ()

'If you can find it, anyone else can find it' — the homework's one-line justification ()

▶ Watch this taught:

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

Why is the frontend-keys lock the one AI has made urgent?

Key-harvesting used to need manual source-reading; an agentic browser does it exhaustively on request — every rendered page, every route, in minutes.

05

Commit → build → release → watch → rollback: the deployment loop

'This is not a one dramatic launch day. You can deploy it over and over and over again.'

The loop, demoed end to end on a live bidding app: push to GitHub; Vercel imports the frontend repo (auto-detects Next.js), env vars pasted in; Render creates a web service from the backend repo (free plan), env vars pasted; the two connect through ONE env value — the Render API URL pasted into Vercel's environment ('that's how it got connected'); logs on both sides are the monitoring ('logs are very important'); errors screenshot straight into the coding agent ('give it to your coding agent in order to understand what went wrong'). DNS closes it: Cloudflare CNAME + TXT records, proxy deliberately OFF ('I don't want Cloudflare's extra security on top of' Vercel's), SSL auto-generates, demo-app.outskill.com live on camera.

The vocabulary is planted by contrast: this IS continuous deployment — versus the old world of 'go to the machine, SSH into it... or writing some YAML script.' CI = merging old and new code; CD = the merged code building and shipping. Rollback is the safety: Vercel's instant-rollback button demoed; databases DON'T roll back unless you keep dev/staging/prod environments. Mobile deploys differ (app stores, bundle IDs, 24-30h review).

Worked example · from the session

The wiring moment most rooms never see: one URL copied from Render, pasted into a Vercel env var, redeploy — and the frontend's error log flips from red to '0 error calls.'

Why it matters

This loop is the session's deliverable — the repeatable path from working-on-laptop to live-on-domain that survives mistakes via rollback.

People get this wrong

Deployment is a launch-day event with a downtime window.

It's a loop that runs many times a day with zero downtime — the platforms build, release, and hold every previous version for instant rollback.

The deployment loop Commit save the state Build CI checks it Release it goes live Watch logs + errors Rollback one click back Ship cheap, watch honestly, roll back without shame — the loop that makes shipping safe under $10
Commit, build, release, watch, rollback: the loop that makes shipping safe
This is not a one dramatic launch day. You can deploy it over and over and over it again.
For your projects

The Cloudflare-proxy-off-with-Vercel detail is a TOC-standards candidate — it's exactly the class of silent double-security conflict that playbook collects.

Go deeper

In one line: Loop: commit (GitHub) → build (Vercel/Render auto) → release → watch (logs both sides; screenshots to the agent) → rollback (instant on Vercel; DBs need dev/staging/prod). Wiring: backend URL as a frontend env var. DNS: CNAME+TXT, Cloudflare proxy off, auto-SSL. CI=merge, CD=build+ship.

Env vars are the connective tissue — and the thing the demo twice paused to fix ()

'If you see any error... take a screenshot and give it to your coding agent' — debugging without helplessness, the session's stated goal ()

Cloudflare proxy off with Vercel: two security layers fight; let the host terminate SSL ()

'You cannot roll back MongoDB until or unless you are keeping 3 different environments' ()

Deployment MCPs (AWS/GCP/Render) let the agent run this loop — 'it will do wonders for you' ()

Certificates require attending BOTH sprint days; recording/transcript/post-reads post 'by tonight' to the resources section ()

▶ Watch this taught:

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

Why does his stack roll back in one click while his database doesn't?

Code releases are immutable versions the host keeps; data mutates in place — only separate dev/staging/prod environments give the database an equivalent of 'the previous version.'

06

Ship cheap, then show: the <$10 receipt and the cohort showcase

He makes the room guess what the lakh-a-day platform costs to run. Guesses hit $2,000. Answer: 'less than 10 dollars.'

The economics close the argument: getout.ai's entire traffic runs UNDER $10/month — because the stack was chosen for it (Golang where it counts, boring managed hosting everywhere else). The honest verdict for the room: Vercel 'without any second thought' for fast validation, Render for the split backend, managed DBs (Atlas/Supabase/Turso) — and 'no need for Kubernetes and all these things... don't go too technical on the engineering front. Just deploy it and see the customer's reaction.' Upgrade SIGNALS, not upgrade schedules: when real money arrives, add real monitoring; when platforms alert you about load, listen.

Then the showcase makes it flesh — ~45 minutes of cohort builds: Role Compass (role-specific AI career intel, 430 roles, weekly refresh); Final Take (AI UGC video studio with director/scriptwriter agents — 100 users hit that day); a real-estate buy-box scraper in progress; Ultron (a local voice assistant, phase 4 of 18); PolarMirror (an AEO/GEO audit platform — how ChatGPT/Perplexity/Gemini/Google AI Overviews see a brand vs 5 competitors, 3-tier reports with 30/60/90-day fix plans — built on Vercel+Supabase over a weekend of Fable credits); and Elodos (peer-to-peer traveler parcel/companion platform, 230+ registrations, live payments). The host's cadence promise: 4-5 stories every sprint.

Worked example · from the session

PolarMirror's origin as told: a cohort member saw Fable credits about to expire, built the whole platform over a weekend, wired payments, and demos the competitor-benchmark report live — the sprint's thesis (ship cheap, ship now) executed by a student.

Why it matters

The showcase is the sprint's evidence layer — six ordinary members shipping is the argument no slide makes; and one of the six is directly relevant to this KB's SEO/AEO priority.

People get this wrong

Real traffic requires enterprise spend.

A lakh a day fits in $10 of boring managed hosting when the stack was chosen for it — spend follows signals, not fear.

Just guess the cost of this project... It's less than 10 dollars.
Don't go too technical on the engineering front. Just deploy it and see the customer's reaction.
For your projects

The under-$10 receipt is a useful counterweight in any hosting-cost conversation — and PolarMirror goes on the OI-074 watchlist.

Go deeper

In one line: Cost doctrine: boring managed stack ≈ <$10/mo at 100k/day; no Kubernetes/Docker for MVPs; upgrade on signals (revenue → monitoring). Showcase as pedagogy: member builds demoed live each sprint — this one surfaced PolarMirror (AEO/GEO audits), Final Take, Role Compass, Elodos, Ultron, and a buy-box scraper.

'Just guess the cost of this project... less than 10 dollars' ()

'Kubernetes, no need to use Kubernetes at all for systems like this' ()

PolarMirror runs the same audit categories as Paul's OI-074 material: brand visibility across ChatGPT/Perplexity/Gemini/AI Overviews + fix plans ()

Final Take hit its first 100 users the day of the session; Elodos reports 230+ registrations — small real numbers, honestly told (2:52:39, 3:18:56)

Community mechanics: recordings + transcripts + post-reads land in 'AI Masterclass & Sprint' resources; both days required for the certificate ()

▶ Watch this taught:

Check yourself

Answer from memory first — the recall attempt is what makes it stick. Then reveal.

What are his upgrade signals, and what aren't they?

Signals: real revenue (add real monitoring), platform load alerts. Not signals: ambition, follower counts, or the existence of Kubernetes.

Every concept, three clicks deep

The same concepts as a quick reference: the closed row is the glance, open is the study card, and every timestamp jumps into the recording.

01Think for scale, ship for ten: the nobody-uses-it-yet correctionSeparate scale problems (traffic) from adoption problems (GTM);›

Separate scale problems (traffic) from adoption problems (GTM); architect scale-compatible via one prompt line naming the deployment stack + a real-user number; ship the boring managed option now. 10 users → anything; 10k → managed hosting + real DB + monitoring; 1M → serious engineering, later.

'You don't have a scale problem. You have a nobody uses it yet problem... opposite problems, opposite decisions' ()

The one-line prompt addition pinning stack + deployment target — 'be specific about the deployment as well' ()

'Pick any of the boring managed option. Ship it. Don't think too much' ()

The $17k leaderboard site as anti-example: virality without architecture — 'this platform will be down' ()

'The gap is not the code' — twice; the agent ('the most advanced intern') wrote fine code; the gap is everything after ()

02No best stack: comfort, agent-compatibility, and the hiring poolStack selection = comfort × agent-compatibility × hiring pool.›

Stack selection = comfort × agent-compatibility × hiring pool. Default: Next.js + Node.js in separate services (Vercel + Render), DB by data shape (Atlas/Postgres/SQLite) + Redis for speed; Golang for high-traffic infra; enterprise Java/.NET only with budget; React Native/Flutter for mobile, native (Kotlin/Swift) for hardware access.

'There is no best tech stack... whatever you are comfortable is the best tech stack for you' ()

The Vue scar: 'really hard to find a developer... I had to migrate from Vue.js to React again, the entire system' ()

Loose coupling = one-click scaling: 'I just need to pay more in order to get more room' on Render ()

Redis's promotion: 'Postgres plus Redis... can handle hundred million users' traffic very easily' ()

Stick to ONE stack per project — agent confusion is a real cost ()

Vercel favored partly because 'Next.js is basically designed via our team only' — as heard; Vercel is Next.js's maker ()

03Git the system, GitHub the shelf — and .gitignore the lockGit = local version-control system (free);›

Git = local version-control system (free); GitHub/GitLab/Bitbucket = hosting shelves monetizing extras. Working loop: init → add → commit → push. .gitignore keeps env files/secrets/node_modules off the shelf — agent-addable ('add git ignore file in both folders'), verified before the first push.

'Git is a process... GitHub is a shelf where you host your code' ()

All three shelves host free; 'these platforms are making money from those extra services' ()

The .gitignore protects against your OWN agent: 'if I make my code public by mistake or my agent have done that' ()

Agent-driven pushing works too: 'I can ask directly agent to do the end to end thing' — the terminal is his preference, not a requirement ()

04The five locks + two prompt lines: security for vibe-coded appsLocks: secrets out of repos;›

Locks: secrets out of repos; keys out of frontends; DB credentials hidden + strong passwords; unguessable admin routes; rate limits on endpoints. Prompt lines: unit tests per query (incl. SQL injection) + no secrets on public routes. Agent audit: 'audit the entire code base — top risks.' Homework: dev-tools key-hunt on your own live site.

AI browsers as attack scouts: give ChatGPT Atlas/Comet-class browsers a harvest prompt and they'll walk every route ()

'Most of the hackers will come and see /admin... keep the URL so that no one can guess' ()

Rate limiting framed as cost defense: an unlimited endpoint is a DDoS invitation ()

The live self-audit caught a real bug: 'a partial refund... reported revenue' mismatch, fixed in the webhook ()

'If you can find it, anyone else can find it' — the homework's one-line justification ()

05Commit → build → release → watch → rollback: the deployment loopLoop: commit (GitHub) → build (Vercel/Render auto) → release → watch (logs both sides;›

Loop: commit (GitHub) → build (Vercel/Render auto) → release → watch (logs both sides; screenshots to the agent) → rollback (instant on Vercel; DBs need dev/staging/prod). Wiring: backend URL as a frontend env var. DNS: CNAME+TXT, Cloudflare proxy off, auto-SSL. CI=merge, CD=build+ship.

Env vars are the connective tissue — and the thing the demo twice paused to fix ()

'If you see any error... take a screenshot and give it to your coding agent' — debugging without helplessness, the session's stated goal ()

Cloudflare proxy off with Vercel: two security layers fight; let the host terminate SSL ()

'You cannot roll back MongoDB until or unless you are keeping 3 different environments' ()

Deployment MCPs (AWS/GCP/Render) let the agent run this loop — 'it will do wonders for you' ()

Certificates require attending BOTH sprint days; recording/transcript/post-reads post 'by tonight' to the resources section ()

06Ship cheap, then show: the <$10 receipt and the cohort showcaseCost doctrine: boring managed stack ≈ <$10/mo at 100k/day;›

Cost doctrine: boring managed stack ≈ <$10/mo at 100k/day; no Kubernetes/Docker for MVPs; upgrade on signals (revenue → monitoring). Showcase as pedagogy: member builds demoed live each sprint — this one surfaced PolarMirror (AEO/GEO audits), Final Take, Role Compass, Elodos, Ultron, and a buy-box scraper.

'Just guess the cost of this project... less than 10 dollars' ()

'Kubernetes, no need to use Kubernetes at all for systems like this' ()

PolarMirror runs the same audit categories as Paul's OI-074 material: brand visibility across ChatGPT/Perplexity/Gemini/AI Overviews + fix plans ()

Final Take hit its first 100 users the day of the session; Elodos reports 230+ registrations — small real numbers, honestly told (2:52:39, 3:18:56)

Community mechanics: recordings + transcripts + post-reads land in 'AI Masterclass & Sprint' resources; both days required for the certificate ()

Tools referenced

ToolCoverageMomentContext
VerceldemonstratedRepo import, env vars, logs, instant rollback, domain + auto-SSL — the session's frontend spine
GitHubdemonstratedTwo repos created live; the shelf both hosts deploy from
Next.jsdemonstratedDefault frontend; auto-detected by Vercel at import
CursordemonstratedHis coding agent for the build + the audit prompt
SupabasementionedNamed as the Postgres-managed alternative; 'higher, costlier side' vs Atlas in his view
Better AuthmentionedWith Clerk, his 'most stable' third-party auth picks; 20+ providers
DockermentionedExplicitly NOT needed for MVPs — with Kubernetes, deferred to money-in-pocket scale

Action items

Resources mentioned

Resources
  • docSession slides + step-by-step deployment documentation ('I'll share the documentation, which covers whatever I have shown you, end to end')
  • docRecording + transcript + post-reads in the 'AI Masterclass & Sprint' resources section ('by tonight')
  • docgetout.ai — his Golang short-link platform with public live stats (uptime, links, traffic)
  • doc1UID.dev — Outskill's in-house auth product ('check it out... let me know the feedback')
  • docCohort builds shown: PolarMirror (AEO/GEO audits — 'polar and mirror dot com' as heard), Final Take (AI video), Role Compass, Elodos (lods.com as heard), Ultron (local assistant)
  • docHomework: the dev-tools secret-hunt on your own live site + optional AI-browser self-audit

Extraction notes

This page was built from an auto-generated transcript, which garbles product and people's names. Those were corrected silently in everything above and logged here for transparency. The warnings flag claims that were true on the recording day but change fast.

Transcript corrections applied

The transcript saysThe trainer actually means
white coding / wipe codingvibe coding
Suga Malya / Sugam / Sagam / Sukam / Sugumthe mentor's name (probably Sugam Malya — UNVERIFIED)
Sumida / Subeda / Sunita / SmiraSumedha (host, per the event page)
Versal / Wersl / Wersal / Wartsel / Wartsail / Watson / WhatsApp (deployment contexts)Vercel
SuperBase / super baseSupabase
get out dot a I / getoutgetout.ai (his short-link platform)
Outskill dot log / outscale dot logthe pay-for-leaderboard site shown (name as heard, unverified)
1 lakh100,000 (Indian numbering)
112 MB free instance512 MB (MongoDB Atlas free tier — corrected)
ChargeGPT / Nchajibiti / JetGPTChatGPT
clogged keysClaude keys
g 2 crowdG2 (review platform)
Shift / shift UISwift / SwiftUI (iOS)
Portland (Android-only language)likely Kotlin (as-heard, unverified)
e value (Render services list)key value (store)
Jammer(unclear — context: Angular-class frameworks; unverified)
TorsoTurso (SQLite cloud database)
Rankboard / brand boardthe demo app's database name (as heard)
Apple Play StoreApple App Store
AI evils and reliabilityAI Evals & Reliability (Day 2's topic)
VC moviesMCU movies (the Avengers — Ultron)
l o d s dot com / yellow dot com / Elodos / Udasthe traveler-parcel platform's domain (as heard, unverified)
polar and mirror dot com / PolarMirror / portal mirrorpolarmirror.com (as heard, unverified)

True on recording day — verify before relying