A Workflow for Grant Discovery and Proposal Development Using Large Language Models: Development and Formative Evaluation

Imagine you are trying to find a hidden treasure chest in a massive, chaotic library where the books keep changing their titles, the maps are drawn in invisible ink, and the rules for who can enter the vault change every day. This is the daily reality for small health-tech companies in developing countries. They have brilliant ideas to save lives using artificial intelligence, but they need money (grants) to prove their tools work. The problem? Finding that money is like searching for a needle in a haystack while the haystack is on fire. They have to scan thousands of funding sources, figure out if they are even allowed to apply, and write complex stories to convince strangers to give them cash. If they spend all their time just looking for the needle, they never get to build the machine that saves lives.

Enter the "Large Language Model" (LLM). Think of an LLM as a super-fast, super-reading robot assistant that can read millions of pages in seconds and write drafts of letters instantly. But here's the catch: if you let the robot run wild, it might make things up, lie about the rules, or accidentally tell a secret. So, the big question isn't "Can the robot do the work?" but "Can we build a cage around the robot so it does the boring work fast, while a human stays in charge to make sure it doesn't mess up?" This paper explores exactly that: building a smart, safe system to help a health-tech team find money without losing their minds or their integrity.

---

The Robot Librarian and the Human Captain

Meet Arkangel AI, a small team in Colombia building smart tools for hospitals. They needed money to test their inventions, but their team was tiny. They couldn't afford a huge army of grant writers. So, they built a "workflow"—a step-by-step assembly line—to hunt for grants. But instead of just humans doing it, they invited a Large Language Model (the Robot Librarian) to help, while keeping a Human Captain (the researchers) firmly at the helm.

Here is how their new system works, broken down into four fun phases:

Phase 1: The Treasure Hunt (Discovery)
In the old days, the team would just type "health AI grant" into a search engine and hope for the best. It was like fishing with a single hook in the ocean. With the new system, the Robot Librarian goes fishing with a net made of 50 different search strategies. It doesn't just look for "grants"; it looks for "innovation prizes," "government programs," and "partnerships" based on the team's specific location and products.

• The Result: The robot scanned 741 different sources in just one week. It found 352 potential candidates. But wait! The robot is fast, not perfect. It also found 220 duplicates (the same grant appearing twice). After filtering out the duplicates and irrelevant results, 83 candidates remained that passed the initial filter and were considered strong, viable options. The human Captain had to step in and say, "Yes, this one is real," before moving it to the next stage.

Phase 2: The Bouncer at the Door (Triage)
Now that the robot has a pile of potential grants, the team needs to decide which ones are worth the effort. This is where the "Triage" happens. The robot checks the rules: Is this for non-profits only? Do we need a partner in Europe? Is the deadline yesterday?
The system uses a special "Do Not Resurface" list. If the robot realizes a grant is for a European company only, it marks it as "Do Not Resurface." This is like a bouncer at a club who remembers, "Hey, you tried to get in last week, but you didn't have the right ID. Don't bother trying again." This stops the team from wasting time on impossible tasks. The human Captain makes the final call on whether to keep an opportunity or throw it in the trash.

Phase 3: The Scoreboard (Win Readiness)
Not all grants are created equal. Some are easy to win; some are like winning the lottery. The team created a "Win Readiness Score" to rank them.

• The Math: They didn't try to guess the exact odds of winning (because they didn't have enough past data to be sure). Instead, they gave each grant a score based on: Are we eligible? Does our mission match theirs? How many people usually win this?
• The Outcome: Out of 49 grants they scored, 5 were so promising that they made it to the interview or finalist stage. The score helped them decide which grants to write about first, rather than guessing.

Phase 4: The Drafting Station (Proposal Development)
This is where the robot really shines. Once the Captain says, "Let's write this one," the robot starts drafting. It writes the boring parts, summarizes the rules, and checks that the citations (the sources of facts) are real.

• The Safety Net: The robot is not allowed to submit anything. Before a single word is sent to a funder, the Human Captain must review it. They check: Did the robot make up a fact? Is there a secret we shouldn't share? Is the story true?
• The Memory: Every single draft, email, and decision is saved in a digital "time capsule" (a version-controlled repository). If they need to apply for a similar grant next year, they can dig up the old work instead of starting from scratch.

What This Actually Means (and What It Doesn't)

The paper is very clear about what they found and what they didn't.

• They Found: The system works! It made the team much faster at finding and organizing grants. They went from finding about 5 or 6 good grants a day to finding about 18 suitable grants per scouting cycle after they tweaked their search strategy. They processed 200+ opportunities and created 46 new, actionable items in just a few weeks.
• They Did NOT Find: They did not prove that using the robot makes them win more money. They only proved that the robot helps them do the work of finding and writing grants more efficiently. Winning a grant depends on many other things (like how good the idea is or how many other people applied), which this system doesn't control.
• The "No" List: The paper explicitly says: Do not let the robot write the whole thing alone. If you let the robot run without a human checking it, you risk lying, breaking rules, or losing your reputation. The robot is a tool, not a replacement for the scientist.

The Takeaway for Curious Minds

This paper is like a blueprint for a "Human-Robot Team" in a world where money is hard to find. It shows that you can use super-fast AI to do the heavy lifting of searching and drafting, but you must keep a human in the driver's seat to check the map and steer the ship.

For small teams in places like Colombia (and anywhere else with limited resources), this workflow is a game-changer. It turns a chaotic, overwhelming mess of paperwork into a clean, organized pipeline. It doesn't guarantee you'll get the money, but it guarantees you won't miss a single opportunity because you were too tired to look. And most importantly, it proves that when you mix human judgment with robot speed, you can build something that is both fast and safe.