Plentiful

What it is, plainly

If you cannot afford groceries, you go to a food pantry, a place that gives food away for free. Historically that meant showing up early and standing in a line for hours, in public, hoping there was still food when you reached the front.

Plentiful is how you get an appointment instead. You find a nearby pantry, book a time, and walk in at that time. That is the whole idea, and everything below is the machinery under it.

Three products, one system

That third one is the part people underestimate. It is production messaging infrastructure at real volume, it carries regulated consent and opt-out obligations, and it exists because a product that only works on a smartphone would systematically exclude the people it is for.

Zero

In 2017 Plentiful was a booking tool at three pantries serving about 600 people, and I was its first full-time employee. The title was Client Adoption Lead, which is a polite way of saying that nobody was using it yet and that was my problem.

So the job was not roadmap work. It was standing in pantries, watching the line, and finding out what the software actually had to be. The thing I learned there is the thing I have used ever since: the constraint is almost never the feature. People did not skip the app because it lacked functionality. They skipped it because standing in a line where everyone can see you is its own cost, and any system that ignored that would fail regardless of how good the scheduling was.

That is where the appointment model came from: not from a roadmap, from the line.

One

Then COVID.

Appointments stopped being a convenience and became the way New York City managed its emergency food lines. Message volume rose 474%. One site checked in 1,700 families in a single day on software that had been designed for three pantries.

That is the part of this I am proudest of, and it is not a growth number. A tool built small delivered big, under load, during the worst months, without a rewrite. The architecture held because the model was right; the model was right because it came out of watching a line in 2017, not out of a planning session.

I was Director of DevOps by then, and still the only person holding product, which I was for the whole nine and a half years. I ran the platform and the operations around it: a four-person field team across all five boroughs, stakeholders spanning funders, two of The City’s largest food nonprofits, and three government entities including the Mayor’s Office of Food Policy.

I also made the call that fixed the unit economics. SMS cost 3.8 cents per transaction against less than a three-thousandth of a cent on web and mobile. At the volumes we were hitting, that gap was thousands of dollars a month, real money for a nonprofit. The migration to web and mobile is what made the scale affordable.

Many

Help Kitchen acquired Plentiful from City Harvest and United Way of NYC in February 2023. I became Co-Founder and Product Director of the company that resulted.

Today: 21 million visits, 1.3 million people, 3,000+ provider locations, live in 15 states in nine languages, with federally regulated eligibility running in five of them. Across all of it, before and after the acquisition, I was the only product lead the product had from 2017 to 2026.

The product operating system: A 290-issue engineering backlog, authored personally, median 35 days to close. RICE scoring as automation rather than as a spreadsheet: an issue form captures reach, impact, confidence, and effort as narrative prompts, framed in writing as “a shared, negotiable reference the team can talk through,” not a formula, and an hourly job writes computed scores back into GitHub Projects. A teammate was still extending it fourteen months later.

One customer call in March 2026 decomposed into 23 scoped tickets in a single day, each citing the same dated meeting as its source.

Regulated delivery: TEFAP has different eligibility rules in every state and they change on statutory schedules. I owned that across five (New York, Colorado, Texas, New Jersey, Montana), and the implementation is approved by the state in all five. The first implementation was a grab-bag of per-state special cases; I replaced it with one declarative JSON Schema per state program, rendered by dual web and mobile renderers, a config-driven architecture generalizable to any state program. One New Jersey deadline was statutory: a senior-age threshold that had to be live October 1st. It shipped behind a feature flag with a fail-safe default in case the flag service itself was unavailable.

The org transition: An eight-year agency relationship wound down and I built the in-house engineering function that replaced it: the CI/CD from zero, the containerized development environments, the first end-to-end test suite. There is a CTO now who out-committed me on the shared codebase, which is exactly right; my 2026 work was product and architecture, not raw commit volume.

What I drew before I could build it

I designed the product. That is not a euphemism for attending meetings about it.

I wrote the flow chart and every line of copy for the SMS system, in English and Spanish, each message written against the character limit so a step never split into two texts on a prepaid phone. I did the rough sketches that became the designs for the day-to-day check-in and messaging screens, which are the most-used parts of the pantry side of the product. Other people typed the code at the time. I have since gone back into most of it myself.

Tooling built around the product

A reporting bot whose interface is a GitHub issue: A stakeholder opens an issue, describes the question in plain English, picks a time range and a breakdown from dropdowns, and adds a label. A workflow authenticates to AWS, runs a Claude Code skill I wrote, and posts a branded PDF back into the thread. Follow-ups are a comment mentioning the bot by name. Underneath, a planner searches a Postgres/pgvector catalog of 8 public datasets, 24 tables, and 1,097 columns, and executes SQL behind a pre-execution linter. A 32-scenario eval suite with deliberately hostile entries exercises it. The part I am proudest of is the interface choice. Nobody had to learn a tool. They already knew how to open an issue.

A conference display, built in days and iterated in public: For the Feeding America Leadership Conference we needed something on a screen that conveyed the scale of the operation to people walking past a booth. Six auto-rotating screens off live data, a geographic heatmap of traffic built from load-balancer logs, and a live feed of survey responses coming in from the floor. Whipped up on a short timeline, then tuned repeatedly against what actually made people stop.

Backlog hygiene as a tool: A small utility that exports a repository’s issues to CSV and applies decisions in bulk (open, close, not planned) with a reason comment attached to every action, so a 290-issue backlog can be triaged honestly in an afternoon and nobody finds a silent close six months later.

What I built with my own hands

I was not a career engineer. I have been technical my whole life (sysadmin, homelab, the person who reads the schematic), but I did not write production code until AI made that reachable. I was early to it and I have not stopped.

What that produced here: the platform’s first Athena data lake and Metabase reporting layer, privacy-engineered so that personal data never leaves the operational database. The Claude-powered reporting system described above. A self-service check-in kiosk, solo, first release in under a week and feature-complete in two, on roughly $150 point-of-sale hardware with $6/month cellular.

And a 2,012-line RFC for the platform’s job-queue migration, against which I ran an adversarial AI review of my own draft and then enumerated in the pull request thread the 13 correctness defects it found: a FIFO ordering bug, a self-contradictory fail-fast spec, a poison-message path that silently burned its own retry budget. Then closed seven open questions with the CTO before any code was written.

That is the shape of it: I can see how the machine works, and the tooling finally caught up to let me build the thing I was looking at.