imaga

The internal tool a team needs most is usually the one IT never gets to.

Too small for the roadmap. Too specific to outsource. Too useful to keep doing by hand.

A client's sales team had exactly this with their own case archive.

Years of projects lived in people's heads, in a 250MB file, and in a slide deck you could only search with Ctrl-F.

Every new proposal started with the same question: have we done something like this before, and where is it?

So two of them — managers, no engineering background — built the fix themselves.

A two-day hackathon, a few weeks of mentoring. The result is a search service: ask a question, get a short answer and cards of the relevant past projects, links included.

Underneath, a vector index of the archive and an LLM to clean up the phrasing.

The first version held together with workarounds. The clean idea — auto-syncing the archive from the slide deck — never worked.

So one of them packaged the data from a presentation overnight and shipped that instead.

This is the work that never reaches the IT backlog. When a manager can build the first working version themselves, it doesn't wait three to six months for a spec-to-release cycle.

That's what our vibe-coding program is for: managers without a technical background build their own internal tools, while IT sets the guardrails — approved data sources, repositories, review.

Not shadow IT. Not replacing engineers. Closing the gap they were never going to get to.

What's sitting in your backlog right now, waiting for IT bandwidth that isn't coming?

Every team has a process that "everyone knows how to do" — and nobody questions how long it takes.

For one crop science company, that process was reading scientific trial PDFs.

Analysts on the agronomy and regulatory teams pulled structured fields out of each paper by hand — Population, Exposure, Comparator, Outcome.

One article took two hours. They handle hundreds a year.

We built an AI assistant to extract those fields. But the part that mattered most wasn't the model.

Before writing any code, we had to lock the extraction schema. That surfaced disagreements between research leads about what even counts as a "correct" extraction.

Resolving that — agreeing on the schema and the evaluation scheme in week one — was the highest-leverage hour of the whole project.

Skip it, and you build a fast pipeline that produces answers nobody trusts.

The second non-obvious decision was the retrieval. A single field scatters across methods, results, and supplementary material.

So instead of fetching the three best passages, we widened the candidate set and let the model filter it down.

Missing a fact costs more than reviewing an extra candidate. Recall on per-field extraction climbed into the 90s with no precision loss we could measure.

And the analyst stayed in the loop by design. The interface puts each extracted fact next to the source passage it came from, so confirming or fixing it takes seconds.

The results:
1) Two hours per article dropped to fifteen minutes.
2) Over 90% accuracy.
3) Roughly 875 analyst hours saved a year.
4) Pilot delivered in seven weeks.

The AI didn't make the team smarter. It moved them from doing the work to checking it — and that only worked because everyone agreed on what "right" looked like before the first line of code.

What's the manual process on your team that everyone just accepts takes hours?

Six months ago she called LLMs “a junior that tells you to jump off a bridge”. Last week she built a three-day report in six hours.

A client asked one of our financial analysts for a conversion funnel built on their CRM data. They wouldn't hand over the CRM data, and they needed the funnel the next day.

She fed the model anonymized dashboard screenshots and asked for the funnel, the hypotheses, and the visuals.

It proposed — she judged. Half the hypotheses got cut by a human who knew the business.

The report took six hours and under a dollar in tokens, against three days of spreadsheet work.

The speed came from the model clearing the routine. A human still decided what was true and caught the stray characters it slipped into the output.

That's why Imaga keeps a human in the loop.

Forward this to whoever on your team is still doing the three-day report by hand.

Six months ago she called LLMs "a junior that tells you to jump off a bridge." Last week she built a three-day report in six hours.

She's a financial analyst at a client company — not a developer.

We'd just run our vibe-coding factory with her team: a two-day hackathon, then mentorship, teaching non-technical people to build their own tools.

Then a real task landed on her: a conversion funnel on the company's CRM data, due the next day, with no access to the CRM data itself.

Analyst fed the model anonymized dashboard screenshots and asked for the funnel, the hypotheses, and the visuals.

It proposed — she judged. Half the hypotheses got cut by a human who knew the business.

Six hours and under a dollar in tokens, against three days of spreadsheet work. The model cleared the routine; she still decided what was true and caught the stray characters it slipped into the output.

The win was hers, not ours.

That's the point of the format: people with no technical background ship their own discovery-MVPs, internal automations, dashboards, and micro-agents — with IT setting the guardrails, and anything sensitive starting on anonymized or synthetic data.

The goal of our vibe-coding factory is an internal champion in your team who keeps shipping after we leave.

If your managers are still waiting months for IT to ship the small things — that's what this is for. Tell us where it hurts: [email protected]

branding.imaga.ai won at the 18th Web Excellence Awards 🎉

Our website for branding services won in the Website category, in the Design Agency and Professional Services subcategories.

Together with last year’s CSS Design Awards and Awwwards recognition, that brings branding.imaga.ai to three awards and five gold placements.