Case study · Clementine Curbside, 2021–2024

Curbside composting, redesigned around the resident.

I co-founded and ran a curbside composting hauler in Boulder County. I also owned the resident experience end to end: the research, the design, and the code. This is what we built, and what it changed.

Three phone screens from the Clementine app: login, a resident's next collection and latest weight with environmental impact, and compostable product ordering

The problem wasn't the brochure. It was the user experience.

Residents put a cart at the curb and never heard anything back. They didn't know what went in it, whether they were doing it right, or whether it mattered. Haulers meanwhile were eating the cost of contamination and had no way to tell the household that caused it. In Colorado, entire loads of materials end up in the landfill due to contamination.

Every hauler I looked at had the same gap: the information existed on the truck, and it died there. So I designed Clementine around one idea: for each customer, let's weigh their cart, analyze their material stream, tell the household what we found that same day, and show them their environmental impact.

One resident's week

The whole product is a feedback loop. Here's how it ran for a household in Erie, Colorado.

  1. Monday evening

    A text: tomorrow is compost day, cart at the curb by 8. Nothing to download, nothing to remember.

  2. Tuesday, on the route

    The cart is weighed and logged. Contamination is recorded per household — with a photo when there is any.

  3. Tuesday afternoon

    An impact text: pounds diverted this week, what that's worth in CO₂, and the running total since they started.

  4. Only if something was wrong

    A separate text with the photo, what the item was, and one line on what to do next time. Specific, not scolding. This is the message that moved contamination.

  5. Anytime

    Residents could text back a question, or a photo of an item, and get an answer. A human on our side, a conversational sorting guide on theirs.

  6. In the app

    Their own portal: next collection, weekly and all-time weights across compost, recycling and trash, impact equivalents, how they compare to other Clementine composters, supply ordering, and more.

A text thread: a reminder that tomorrow is recycling day, then a post-collection message with a recycling score and a photo of a plastic bag flagged as an item not to include next time

The feedback text, with photo evidence and one thing to change. This example is from the Bluterra resident pilot in late 2020; Clementine ran the same loop for compost from 2021.

Clementine app screens showing all-time compost, waste and recycling weights, impact equivalents, and a comparison to the average Colorado resident

The composter portal. Every number on it came from a scale on the truck, not an estimate.

What changed

Closing the loop did what education campaigns and cart stickers never had.

<1%contamination by weight across all routes — effectively zero for curbside organics
500+households, grown entirely by word of mouth
$0spent on paid acquisition
2024operations acquired by Wompost
"After every collection, they send me information about how much waste I collected that week and how that impacts the environment, which is really meaningful to me."

The other customer: commercial organics programs

We ran the Town of Erie's own composting program across four town facilities. The same loop applied, at building scale: every collection weighed, contamination scored per location, and a year-end impact report the program manager could put in front of council.

Per-location grades turned an abstract goal into a friendly competition between Town Hall and the Community Center. That's the municipal side of behavior change — the buyer needs configurability and reporting; the person emptying the bin needs to know they're doing it right.

Page from the 2022 Town of Erie compost impact report showing an overall B+ contamination score, tips for improvement, and per-location compost weights and grades

Town of Erie 2022 report: contamination and recommendations by facility.

Then: Bluterra, 2020–2026

Closing the same loop with computer vision

Clementine's feedback loop worked because a person looked at every cart, but that doesn't scale to a municipality. Bluterra was built to put that analysis on every truck: a camera over the hopper, a model that flags contaminants at the tip, reports that go to the household and to the program manager, and a web dashboard to keep a pulse on the health of the waste stream.

Bluterra tip-event screen: a hopper frame with every item boxed and labeled by material (cardboard, plastic film, bagged materials) with confidence scores, plus coverage, recycling percentage, route, frame quality metrics, and a pending review status

One tip event from the Rumpke beta. Every item is detected and labeled by material with a confidence score; the event carries route, recycling content, coverage, and frame-quality metrics, and auto-flagging adds it to a human review queue if needed.

Analysis in production

Hundreds of video files a day, 10-hour continuous streams or remote uploads. Cloud inference with a local YOLO fallback for 99.9% uptime.

Humans stay in the loop

Low-confidence and quality-issue frames are auto-flagged for review; users can flag too. Every correction goes back into the training set.

The Rumpke beta

Rumpke Waste & Recycling asked for a version they could run themselves: multi-user, drag-and-drop batch upload, per-tip results with recycling-content thresholds, CSV export. We beta-tested it on their trucks with both existing camera hardware and our custom-built hardware units.

Two-minute walkthrough of the Rumpke build: upload, analysis parameters, and tip-event results. Open on YouTube if the player doesn't load.

Other analysis projects

Assisted The Recycling Partnership with material analysis of recycling carts to quantify the impact of a program intervention — both before and after intervention.

Bluterra contamination analysis page for The Recycling Partnership: carts analyzed, share with contamination, and items per 100 carts by contaminant type, before and after a program intervention

Residential recycling cart analysis for The Recycling Partnership, pre- and post-intervention.

What I took from six years of this

Lasting change rarely comes from a brochure, especially when materials are constantly changing and people are constantly moving. Instead, personalized and instant feedback makes all the difference, and that is possible today — now more than ever. The hard part of a resident-facing waste product isn't the app; it's making sure the UX is streamlined and that you get the details right, every time.

And AI belongs exactly where a person can't be: on every truck, at every tip, and providing users instant help, 24/7. To build trust, these tools have to be held to a high bar — misidentify a customer or a material in their stream and credibility shatters fast.