Finish an AI Research-to-PDF Document Pipeline

We’d like you to propose a fee, ideally based on a time estimate for the work. If you’re interested, we’re happy to arrange a call so you have enough detail to scope it properly.

We’re a small professional services firm. Part of our process is researching an individual’s public online presence and turning the results into a set of clean, annotated PDF documents that go into a formal submission pack. It’s high-volume, repetitive work that we’re automating.

We’ve built a first version ourselves, but it is to demonstrate the workflow only as it doesn’t work error free (we are not developers). We now need a developer to make it production-grade.

The pipeline

AI research agents search for articles, reviews, listings and coverage about a named subject.

Each relevant source comes back as a JSON object matching a fixed schema. This holds the URL, the source metadata, and the exact phrases that should be highlighted in the final document.

Our team reviews the results, confirms what to keep, rejects what we don’t want, and can send feedback to the agent to rerun the search.

A web app takes the approved sources, renders each URL as a PDF, and applies the annotations to the phrases named in the schema.

Our team can adjust the annotations in the app, or swap in a manually rendered PDF where the automatic one looks poor, then download the finished set.

Where we are now

The research prompts can return usable results and roughly the right JSON. What we don’t have is a setup that runs this in bulk, cleanly and repeatably.

The web app exists. I built v1 myself using an AI app builder and its hosting. I’m not a developer, and the main purpose of it was to demonstrate the workflow rather than to run it properly. It needs an honest review, and possibly replacing.

What we need

1. The research side

A setup that handles bulk research properly.

Right now we’re using a lot of Claude chats to do the research and then either export a CSV or run the highlights manually. It works, but it isn’t something we can onboard a team onto, so we can’t scale. We’ve been trying to build a Clay table for the research step because it keeps the output as structured data, but it isn’t user-freidnly . Better suggestions are welcome.

The review stage is the biggest gap. We need somewhere our team can see the results, confirm the ones we want, reject the rest, and rerun the search with feedback to the agent.

2. The translation problem

The JSON schema includes verbatim phrases from each article, the phrases that led the agent to include that source. When the article is in another language, two things have to be true at once. The review screen has to show our team those phrases in English so they can judge them. And the phrases stored in the schema have to match the text in the rendered PDF exactly, or the app won’t find them to highlight.

So the translation used at the review stage and the translation used at the render stage have to produce identical output. Thoughts on Google, DeepL or another API are welcome.

3. The web app

Review the existing app end to end and tell us what’s solid and what isn’t.

Fix what’s broken and harden the rendering. The aim is a PDF that looks as close to the original page as possible, with the ad clutter stripped out, set into standard margins on A4 so there’s room for notes and annotations. There’s currently no translation step, and every non-English page will need translating at the point it’s rendered.

Make the annotation layer reliable so the phrases from the schema are applied accurately, and our team can adjust them in the interface.

Keep a manual upload path so a badly rendered page can be replaced by hand. We know full automation creates more problems than it solves once you hit an edge case, so we want a human review point where we can step in.

Finish with a clean batch download of the completed set.

4. Hosting

We’d are considering running inside the existing platform (Remy), since the hosting is bundled and it’s simple for a non-technical team to manage. But we’re genuinely open to other solutions. If you think it should be rebuilt or moved somewhere more suitable, please tell us why and what it would cost.