Try it: lookuplater.app Code (MIT): github.com/impactvelocity/lookuplater-app

Turn your questions into a weekly newsletter: text Lou your questions, live your life, read it Friday

Inspiration

Why do onions make you cry? Who invented the stapler? You're out with friends, or travelling, or in the middle of something else. You want to know, just not right now.

There's no good way to look something up later. Searching gets you the answer right then, and one search turns into twenty minutes and a dozen tabs. So you don't search, and the question's gone.

Every question has two moments: the moment you wonder, and the moment you have time to read. Search squashes them together. I built LookupLater to pull them apart.

What it does

You text a question to LookupLater, then put your phone back in your pocket. Type it or dictate it. Nothing texts you back, so you can carry on with your day. Think of it as an inbox for your questions.

After you hit send, three things happen:

  1. NVIDIA Nemotron Nano, running on Nebius Token Factory, reads the question. You probably sent it in a hurry, so it fixes typos, spells out abbreviations and adds context when the question needs it. If your phone hears "why do cats need blankets," it searches for "why do cats knead blankets." Clear questions are left alone.
  2. Tavily searches the web for the cleaned-up question and brings back sources and images.
  3. Nemotron Super, also on Token Factory, writes the answer from those sources and cites each one. If the sources don't have the answer, it says so instead of guessing.

The three steps: Nemotron Nano cleans up the question, Tavily searches the web, Nemotron Super writes the answer

Answers land in your LookupLater inbox as they're ready. Every Friday at 8 am in your timezone, the whole week arrives as one newsletter. Every story in it is one you asked for. You choose whether it carries full answers or a couple of lines each with a link, and whether it includes the whole week or only what you haven't read. No unread answers means no email.

The Friday newsletter email and its options

You can pick one of six themes (Classic, Editorial, Search, Messages, Post and Minimal), then give the newsletter a name, a tagline and your own sign-off. You can link up to five phone numbers, so your partner or your whole family can text into the same Friday issue.

Six newsletter themes: Classic, Editorial, Search, Messages, Post and Minimal

Texting is the main way in, but there are others: the web app, an Apple Shortcut you can run from Siri or use the API.

Five family members text Lou their questions, and one Friday newsletter answers them all

LookupLater is open source. You can use the hosted app, run the full app yourself, or deploy the headless version, a Cloudflare Worker with no UI. Add your Token Factory and Tavily keys, point your SMS provider's webhook at it, and you have your own Friday newsletter.

How I built it

It's a Turborepo monorepo. The main app is Next.js 16 with Tailwind v4, Base UI, Clerk for auth and MongoDB for storage.

Every way of adding a question (SMS, web, API) goes through one function that saves the question and starts a durable workflow with the Vercel Workflow SDK. The workflow runs the three steps as separate retryable steps, up to three attempts each, and writes progress to the lookup's status so the app can show it moving along.

How a text becomes Friday's email: Zernio, Next.js on Vercel, the lookup workflow, the inbox, the weekly workflow and Resend

I call both models through Token Factory's OpenAI-compatible chat API. Nano runs in JSON mode and returns { shouldRewrite, rewritten, reason }. It runs on every question, so it needs to be fast. Super writes Markdown with inline [n] citations and a sources list. Splitting it this way keeps the small model on the quick, frequent call and saves the big one for the writing.

Inbound texts come from Zernio through a webhook. It only acts on incoming messages from numbers linked to an account, and it de-duplicates on the message id.

For the newsletter, a cron job runs every Thursday and starts one workflow per subscriber. Each workflow sleeps until Friday 8 am in that person's timezone, then builds the issue and sends it through Resend. The templates are React Email components in a shared package, so the settings page previews the real email from your own past week.

Challenges I ran into

Nemotron Super is a reasoning model, and at first it could spend its whole token budget thinking before it wrote the answer. Sending chat_template_kwargs: { enable_thinking: false } with answer requests fixed it.

Super also sometimes emits citation tokens like 【1†L0:16-L0:47】 instead of [1]. I wrote a small sanitizer that rewrites them before the answer is saved.

Email had its own problems. Gmail clips any message over about 102 KB, and a full answer is around 20 KB, so a busy week would get cut off. The renderer now lowers the number of full answers until the HTML is 90 KB or less. Gmail also ignores object-fit, so images are served as signed 360×240 crops. Dark mode needed a prefers-color-scheme block that matches the inline colours with attribute selectors, which means every colour in the palette has to be unique.

Accomplishments that I'm proud of

It's a complete product, not a proof of concept. You can sign up, link your number, text a question and get a real Friday newsletter.

The rewrite step catches dictation mistakes that would otherwise send the search somewhere useless, and the app shows what it searched for so you can see the fix.

Answers cite their sources inline, and when the sources don't cover the question, the answer says so.

Anyone can run their own copy. The headless Worker needs a Token Factory key, a Tavily key and an SMS webhook.

What I learned

Matching the model to the job matters. A fast, small model is right for a check that runs on every message, and the larger model earns its cost when it's writing something people will read.

Reasoning models need explicit settings. With thinking left on, Super could use up its budget before writing anything.

Email HTML is its own platform with its own limits, and size, image handling and dark mode each needed a different workaround.

Durable workflows suit a product that waits. "Sleep until Friday 8 am in this person's timezone" is one step in a workflow instead of a queue I have to manage.

What's next for LookupLater

  • Measure real use: how many questions people send each week and what time of day they send them.
  • Test the headless Worker's SMS parsers against live providers. The parsers for providers other than Zernio and Twilio have only been tried with sample payloads.
  • Move the headless Worker into its own public repository.

Built With

  • cloudflare-d1
  • cloudflare-workers
  • cloudflare-workflows
  • nebius-token-factory
  • nvidia-nemotron
  • pnpm
  • react
  • resend
  • tavily
  • zernio
Share this project:

Updates

Submission history