I build AI tools whose claims you can check. Every project here backs its output with something a human can verify in seconds — a verbatim citation, a rubric a reviewer can read, a signal you can trace — instead of asking you to trust a model's say-so.
Why that framing? The easy version of "AI for nonprofits" is a thin wrapper around a chat model: it's agreeable, it's confident, and in a domain where a wrong "yes" costs an organization a grant or ships a broken laptop to a family, agreeable-and-confident is dangerous. The harder, more useful version separates the model from the decision — the model observes, deterministic code you can audit decides — so the tool's trustworthiness doesn't depend on the model being honest.
I come at this from hard tech: four years of field service on production critical semiconductor equipment in a live fab, keeping high volume production lines in service, now on a Master Electrician track. That background, disciplined engineering against a spec where a wrong call takes down a production line, is where the "verify it in code, not vibes" instinct comes from, and it is the part a no code AI consultant cannot replicate.
Each runs free, client-side, and offline-capable — deployable by a nonprofit at $0/month, no accounts, no cloud, no API bills. Click a demo, or clone and run it locally in ~30 seconds (below).
| Project | What it proves | Live demo | Status |
|---|---|---|---|
| GrantMatch — RFP compliance checker | Extracts every requirement from a funder's RFP with a verbatim citation, then refuses to mark your draft "compliant" unless it can quote your own words back. Catches the model fabricating coverage — live. | Try it → | ✅ Shipped |
| ColdWatch — edge spoilage sensor | A ~$21 fridge/freezer monitor that predicts failure hours early, ignores door-opens so the alarm stays trusted, and alerts offline — with zero AI at runtime (knowing when not to use a model is the skill). Same logic ships in browser JS and ESP32 C firmware. | Try it → | ✅ Shipped |
| ReClaim Vision — e-waste grading | Industrial-grade inspection discipline for donated electronics: the vision model only observes, a versioned rubric with severity floors decides, and no blocking-defect device can ever be graded "refurbish" — enforced against the model and a human override. | Try it → | 🛠️ Demo · QA in progress |
| BenefitsBridge — benefits eligibility Q&A | Answers SNAP and energy-assistance questions only from official rule text it retrieved, quoting the rule for every statement. Ask about a program it hasn't loaded and it refuses and routes you to a caseworker. Any categorical "you qualify" is blocked outright, in English and Spanish. | Try it → | ✅ Shipped |
Live site: skeeter-spec.github.io/edge-ai-portfolio — all four demos, public, no sign-in. Prefer to run it yourself? Clone and serve locally in ~30 seconds (below).
git clone https://github.com/skeeter-spec/edge-ai-portfolio.git
cd edge-ai-portfolio/grantmatch/build # or coldwatch/build, or reclaim-vision/build
python3 -m http.server 8000
# open http://localhost:8000No npm install, no bundler, no dependencies to fetch — the apps are plain HTML/CSS/JS. Each app's
folder has a full case study (case-study.md) and its own README. Each build/ has a test.mjs
you can run with node test.mjs to see the verification logic proven with zero model calls.
- 100% client-side. No servers, no telemetry, no analytics. In the apps' "live" mode the only network call is to a model running on your own machine (local Ollama).
- No API keys, no secrets, no vendor lock-in. Nothing to leak, nothing to bill. A nonprofit can run this forever on hardware it already owns.
- The verification is plain, inspectable code — no AI is required to check the AI.
compliance.js,rubric.js+triage.js,anomaly.jsare each a few hundred readable lines, each with automated tests asserting the safety properties (including source-level tests that fail the build if a supposedly-offline module contains a network call). - Demo mode ships pre-computed real model outputs so the demos are instant and reliable, while the deterministic verification core still runs live in your browser — so you're watching the real guarantee, not a canned animation. Point any app at your own local model for the full pipeline.
- Data safety: demos use public or synthetic data only. No client's confidential material appears anywhere in this repo.
Three more edge-AI tools for nonprofits are in the pipeline, same thesis: OffGrid Assistant (a fully offline LLM assistant on a Raspberry Pi), ShiftPilot (a volunteer-scheduling agent whose tool order is decided by deterministic code, never by the model), and WattWise (energy-retrofit ROI from sensor data). This repo will grow as each ships.
Built by Keaton Taliaferro. Six years in hard tech: two years characterizing custom electronics at board level, then four years on production critical semiconductor equipment in a live fab, where I led the doping design of experiments program that became process engineering's reference going forward. Now on a Master Electrician track, building AI tools for nonprofits.
- Email: scholarkeaton@protonmail.com
- GitHub: @skeeter-spec
- LinkedIn: Keaton Taliaferro
GNU AGPL-3.0 · Copyright © 2026 Keaton Taliaferro.
Read it, run it, learn from it. But if you build on this code or deploy a modified version — including as a hosted service — the AGPL requires you to release your source under the same license. In short: it stays open; no one gets to take it closed-source. Want to use it under different terms? I hold the copyright — get in touch.