https://www.putthrough.com/blog Field notes from PutThrough on building web apps, AI agents and mobile apps in a week, and on taking AI-built apps safely to production. en-us Mon, 05 Oct 2026 09:00:00 GMT https://www.putthrough.com/blog/ai-customer-service-agents https://www.putthrough.com/blog/ai-customer-service-agents Mon, 05 Oct 2026 09:00:00 GMT contact@putthrough.com (Anubhav Mehrotra) Building An AI customer service agent answers customers in chat or email from your own help center and data, and hands off to a person when it can’t help. They handle routine, well-documented questions well and struggle with exceptions: Gartner found only 14% of service issues are fully resolved through self-service, and 87% of customers want a way to reach a human. Buying one is priced per resolution, from about $0.50 on HubSpot to $0.99 an outcome on Intercom Fin and $1.50–$2.00 on Zendesk. Building one costs model fees of about $10–$192 a month for 3,000 conversations in our illustration, plus the build. From January 1, 2027, California will require large businesses to offer a human within 15 minutes, and the EU already requires telling people they’re talking to AI. https://www.putthrough.com/blog/ai-receptionist https://www.putthrough.com/blog/ai-receptionist Mon, 05 Oct 2026 09:00:00 GMT contact@putthrough.com (Anubhav Mehrotra) Building An AI receptionist is a voice agent that answers your business phone line, handles common questions from your own FAQs and documents, takes messages and transfers callers to a person. At 1,000 three-minute calls a month, published AI receptionist services cost about $0.50–$1.85 a call, voice platforms you build on about $0.25–$0.66 a call in usage plus the build, and a human answering service about $8 a call. Answering inbound calls with AI isn’t covered by the US robocall rules, but calling people with an AI voice needs their prior consent, many states require every party’s consent to record, and from 2027 California requires large businesses to offer a person. https://www.putthrough.com/blog/custom-crm-vs-hubspot https://www.putthrough.com/blog/custom-crm-vs-hubspot Mon, 05 Oct 2026 09:00:00 GMT contact@putthrough.com (Anubhav Mehrotra) Building Build a custom CRM when your core process doesn’t fit a sales pipeline, or when you pay for many seats but use few features. Buy HubSpot, Pipedrive, Salesforce or Zoho when your process is a standard pipeline or you need marketing tools today. The deciding arithmetic is seats × price × months: ten seats on HubSpot Sales Hub Professional come to about $33,900 over three years at list price, including its required onboarding fee, while a small custom stack costs about $65 a month to run, plus the build. https://www.putthrough.com/blog/how-to-build-an-ai-agent https://www.putthrough.com/blog/how-to-build-an-ai-agent Mon, 05 Oct 2026 09:00:00 GMT contact@putthrough.com (Anubhav Mehrotra) Building An AI agent is software that uses a language model to decide its own next steps (what to look up, which tool to call, when it’s done) to finish a task for you. To build one, pick one job with a measurable outcome, collect real past cases as a test set, give it only the tools it needs, write its instructions from how your team already works, and launch with a person reviewing its work. Model fees are usually the small part: in the illustration below, 2,000 conversations a month cost between about $6 and $128 depending on the model. The real risks are wrong answers and actions that can’t be undone, so every agent needs sources, a handoff to a person and a spending cap. Often, a simple workflow or a single model call does the job better than an agent. https://www.putthrough.com/blog/turn-website-into-app https://www.putthrough.com/blog/turn-website-into-app Wed, 30 Sep 2026 09:00:00 GMT contact@putthrough.com (Anubhav Mehrotra) Building Yes, but not by wrapping it. Apple’s guideline 4.2 rejects apps that are a “repackaged website”. Google Play allows an app that shows a website only with the site owner’s permission, and rejects apps with limited functionality. Four routes work: a progressive web app with no store at all, a Trusted Web Activity for Google Play, Capacitor (your web app plus native features), or a React Native rebuild. Publishing costs $99 a year for Apple and a one-time $25 for Google. https://www.putthrough.com/blog/how-much-does-it-cost-to-build-an-app https://www.putthrough.com/blog/how-much-does-it-cost-to-build-an-app Wed, 30 Sep 2026 09:00:00 GMT contact@putthrough.com (Anubhav Mehrotra) Building Most app projects cost $10,000–$49,999, according to Clutch’s September 2026 pricing guide, which is built from verified client reviews. The average is $90,780, because a few large projects pull it up. What sets your price is scope: how many kinds of users, core workflows and integrations the app has, and whether it runs on the web, iOS or Android. The kind of app matters far less. Running it is a separate bill: a small production stack lists at about $91 a month, plus Apple’s $99 a year and Google’s one-time $25. https://www.putthrough.com/blog/supabase-vs-firebase https://www.putthrough.com/blog/supabase-vs-firebase Wed, 30 Sep 2026 09:00:00 GMT contact@putthrough.com (Anubhav Mehrotra) Building Choose Supabase if your app needs relational data, one set of access rules for every client, a predictable monthly plan and the option to self-host. Choose Firebase for offline-first mobile apps inside Google Cloud with pay-per-use pricing that scales to zero. The old “SQL versus NoSQL” split no longer holds, because Firebase now offers Postgres through SQL Connect. In production, Supabase starts at $25 a month plus compute. Firebase starts at $0 and charges per read and write, and its budget alerts don’t cap the bill. https://www.putthrough.com/blog/what-is-vibe-coding https://www.putthrough.com/blog/what-is-vibe-coding Wed, 30 Sep 2026 09:00:00 GMT contact@putthrough.com (Anubhav Mehrotra) Audit & fix Vibe coding is building software by describing what you want to an AI and accepting the code it writes without reading it. Andrej Karpathy coined the term in February 2025, and Collins made it its 2025 Word of the Year. It is a fast way to build a prototype and a risky way to launch one. In Veracode’s 2026 tests, AI-generated code passed security checks only 56% of the time. The public incidents are mostly missing access rules and exposed keys, and both can be checked before real users arrive. https://www.putthrough.com/blog/supabase-rls-not-working https://www.putthrough.com/blog/supabase-rls-not-working Tue, 29 Sep 2026 09:00:00 GMT contact@putthrough.com (Anubhav Mehrotra) Audit & fix When Supabase row-level security (RLS) seems not to work, the symptom tells you the cause. An empty result with no error means no policy lets that request read the rows, often because it isn’t signed in, so auth.uid() is null. “New row violates row-level security policy” means a WITH CHECK failed, or the insert asked for the row back without a select policy. Updates that change nothing need a select policy as well as an update policy. Data anyone can read means RLS is off, a view bypasses it, or a secret key reached the browser. https://www.putthrough.com/blog/what-fits-in-a-one-week-build https://www.putthrough.com/blog/what-fits-in-a-one-week-build Sat, 26 Sep 2026 09:00:00 GMT contact@putthrough.com (Anubhav Mehrotra) Building Yes, if the MVP is one product with one core workflow. A one-week build fits exactly that, done to production standard: sign-in and roles, payments, an admin view, two or three integrations, and the monitoring, backups and deploy pipeline around them. It does not fit several products at once, heavy real-time or offline work, formal compliance programs, or a large migration out of an old system. The scope is cut and written down on a call before the week starts, so nobody discovers on Thursday that it was never going to fit. https://www.putthrough.com/blog/lovable-demo-to-production-checklist https://www.putthrough.com/blog/lovable-demo-to-production-checklist Sat, 26 Sep 2026 09:00:00 GMT contact@putthrough.com (Anubhav Mehrotra) Audit & fix Before a Lovable app takes real users or real money, check seven things: every Supabase table has row-level security with tested policies; no secret key is in the browser; sign-in, sessions and password reset behave; Stripe webhooks are verified and handled idempotently; errors and downtime alert someone; backups exist and a restore has been tested; and deploys can be rolled back. Most of this also applies to apps built with Bolt, Cursor or Replit. https://www.putthrough.com/blog/ways-ai-built-apps-fail-in-production https://www.putthrough.com/blog/ways-ai-built-apps-fail-in-production Sat, 26 Sep 2026 09:00:00 GMT contact@putthrough.com (Anubhav Mehrotra) Audit & fix AI coding tools are very good at the happy path: one user, one browser, everything working. Production is everything else, and AI-built apps fail there in the same ways again and again — open database access, secrets in the browser, payments that charge without updating anything, no backups, no monitoring, and store rejections for mobile. These are the twenty failure modes we check every AI-built app for, grouped by area, with what each one looks like and what fixes it. https://www.putthrough.com/blog/why-we-dont-bill-by-the-hour https://www.putthrough.com/blog/why-we-dont-bill-by-the-hour Sat, 26 Sep 2026 09:00:00 GMT contact@putthrough.com (Anubhav Mehrotra) How we work We price a build by its scope, not by our hours, because an hourly rate puts all the risk of a slow week on you. The scope is written down and the price fixed before work starts; if we estimated wrong, that is our problem. Scope changes are handled as a new, separate quote, so the original price and launch date hold. Hourly billing is still the better model for open-ended work with no clear finish line.