If I Had One Shelter Partner and Six Months, Here's What I'd Build
I'm a machine learning engineer, and lately I've been staring at a mismatch that's hard to unsee.
The headline numbers in animal welfare are genuinely encouraging. Roughly 5 to 6 million dogs and cats enter U.S. shelters each year, but outcomes have improved dramatically over the past decade: around 4 million animals found homes in 2025, and more than two-thirds of U.S. shelters now clear the 90% save-rate "no-kill" benchmark [1][2].
But progress isn't the same thing as victory. In 2025, approximately 396,000 dogs and cats were still killed in U.S. shelters—the sharpest decline since 2020, yet still a staggering number [1][3].
The more I've learned about shelter operations, the more it seems that a meaningful part of that remaining gap isn't fundamentally a love problem or even a money problem. It's an information problem.
Animals languish because no one flagged them. Adoptions fail because the match was wrong. Foster caregivers quit because they couldn't get help when they needed it. Urgent cases sit in inboxes. Staff spend time hunting for information instead of acting on it.
That's exactly the kind of problem modern machine learning is good at solving: triage, matching, retrieval, and surfacing the signal that's already buried in the data.
So I've been asking myself a simple question:
If I had one shelter partner and six months to work with them, what would I actually build?
Not moonshots. Not research projects. Systems that could realistically improve outcomes for animals and the people trying to help them.
Here are ten ideas, roughly in the order an animal moves through a shelter. Some could be deployed with today's models and a motivated shelter partner. Others would require historical data and careful validation. I've tried to be honest about both the opportunities and the hard parts.
Supporting the people who care for animals
1. A RAG system for fosters. Foster homes are one of the biggest levers in modern sheltering, because they expand a shelter's capacity beyond its kennels—but fosters are volunteers caring for fragile animals at home, and their hardest moments come after hours. A neonatal kitten that won't eat at 2am. A post-surgical dog that looks wrong on a Sunday. The engineering move is a retrieval-augmented assistant grounded only in the shelter's approved materials (foster handbook, vet protocols, care guides), answering with citations.
2. An adopter onboarding assistant. Same retrieval backbone, pointed at new adopters: "my rescue dog has been hiding under the bed for two days—is that normal?" The first weeks are when adoptions are most fragile, and the questions that derail them are usually about behavior surprises and "is this normal?" Good support during week one is a direct lever on the return curve, and it's buildable today on public behavior resources.
3. A multilingual care-info translator. Many shelters serve communities they can't reach in their own materials. A tool that renders care guides and adoption paperwork into accurate, plain-language versions across languages widens the adopter and foster pool. Instant to demo, obvious equity payoff, and a clean showcase of evaluation discipline (you have to verify the translations don't quietly mangle medical instructions).
Getting the right animal into the right home
4. An adoption-matching recommender. Match adopter lifestyle and preferences to the animals likely to actually fit. A recommender can't fix every mismatch, but it can stop sending a high-energy working breed home with someone who wanted a couch companion. Every prevented return is a stressed animal spared a round trip and a kennel freed up for the next one.
5. Return-risk scoring. The mirror image of matching: score each adoption for return risk at the moment it happens, and trigger proactive follow-up for the high-risk ones. Because the riskiest window is so short, even a single well-timed check-in call can change the outcome. This one needs a real org's historical data to train on, but the signal is strong and the intervention is cheap.
Turning raw inputs into data
6. Medical and behavior note structuring. This is the unglamorous one I'd be tempted to build first, because it's load-bearing for half the others. Vet and behavior notes are free text—which means they're effectively invisible to analysis. An NLP pipeline that parses them into structured fields (conditions, behaviors, flags) turns a write-only archive into a dataset. Return-risk, health-trend tracking, and matching all get sharper the moment this exists.
7. Computer-vision body-condition scoring. Photo-based triage at intake: estimate body-condition score, flag visible injury or distress, and push the urgent cases to the front of the medical queue. This is the most technically distinctive build (vision rather than tabular/NLP) and also the most data-hungry, so I'd be upfront that doing it well is harder than doing it at all.
Driving adoptions and funding the work
8. Adoption-listing A/B optimization. Which photo, which bio, which channel actually drives an adoption? Most shelters have no marketing-science layer at all. An experimentation system that tests listing variants and promotes the winners is using data the shelter already generates, and photo quality and listing copy are known, strong levers on time-to-adoption.
9. Donation and grant targeting. A model that predicts which donors—especially lapsed ones—are most likely to give lets a tiny development team spend its hours where they convert. This is bog-standard propensity modeling in any other sector; it's just rarely applied here. More efficient fundraising is more funded lifesaving programs.
10. Incoming-call and email triage. Shelters drown in inbound messages: surrenders, lost-pet reports, adoption questions, donations. A classifier that routes them and surfaces the urgent ones (a found stray, a medical emergency) reclaims staff hours and shortens response times.
Why most shelter AI projects would fail
The biggest risk isn't model quality.
It's workflow fit.
Shelter staff are already stretched thin. A model that achieves impressive benchmark performance but requires people to fundamentally change how they work is unlikely to survive deployment. The most valuable system is usually the one that quietly removes friction from an existing process.
I'd rather deploy a reliable tool that saves five minutes per animal than a state-of-the-art model nobody trusts.
In animal welfare, adoption matters more than sophistication—not just for animals, but for technology too.
Where I'd start
If I had six months and one shelter partner, I wouldn't start with the flashiest model.
I'd start with the systems that create leverage for everything else.
First, I'd build medical and behavior note structuring. Most shelter data is trapped in free text, and turning it into structured information makes every downstream system better.
Second, I'd build a foster-support RAG assistant grounded in approved shelter and veterinary materials. Foster capacity is one of the strongest drivers of lifesaving, and volunteers often need help when staff aren't available.
Third, I'd build an adopter onboarding assistant focused on the first few weeks after adoption, when returns are most likely.
None of those projects require frontier AI. They require careful evaluation, thoughtful deployment, and close collaboration with shelter staff.
That's the broader lesson I've come away with after thinking about this space. The binding constraint in animal welfare isn't model sophistication—it's that the people holding the data and the people who can model it almost never sit at the same table.
A shelter doesn't need all ten of these systems. It needs one that works, deployed and trusted, closing some piece of its lifesaving gap.
If you're a technologist looking for work where the feedback loop is real animals and measurable outcomes, there are few better places to point your skills.
Animal welfare doesn't need breakthrough models nearly as much as it needs thoughtful people willing to build useful things.
I know where I'm pointing mine.
I'm Emma Junqua, a machine learning engineer looking to put these skills to work in animal welfare. If you run a shelter or rescue, work in the space, or are an engineer who wants to build something on this list, I'd love to compare notes—reach me at @emmajunqua.
Sources
[1] Best Friends Animal Society, 2025 National Shelter Data Report
[2] Best Friends Animal Society, Animal Shelter Statistics Dashboard
[3] Best Friends Animal Society, New Data Shows Sharpest Drop in U.S. Animal Shelter Deaths Since 2020