The Hope Group · Boston · Ethical AI lab
AI Tools for Economic Justice.
I'm the Founding Director of Technology + Design at The Hope Group, an ethical AI lab in Boston. I design, develop, and project-manage the lab's software, from the first conversation with the people who will use it to the production deploy.
Fig. 1 · Four beacons, all lit. Move the light, or pick one.
- RoleFounding Director of Technology + DesignI lead technology and design at the lab, including its work on local, open-weight models inside the apps it makes.
- On every buildDesign, development, project managementThe models, the pipelines, the interfaces, and the plain-language explanations that ship with them.
- MethodCo-designed, then tested in real routinesEach tool is built with the people who use it and reshaped by what they tell us.
- Rule of thumbHumans when needed, code when convenient, AI when necessary.A person reviews and decides. The software shows its work.
What I've built
Pragmatic, ethical AI tools for minority-owned small businesses, each built with the people who use it.
Each one started with a specific chore that was eating someone's day: reading procurement portals, retyping purchase orders, chasing bug reports through a spreadsheet.
Each tool is accessible only to the partner who helped make it – it's more theirs than ours, and so the below demos are previews only. For a real demo of our biggest project, Project Lookout, email me at stephen@hopegroup.ai.
Tool · Live · In daily production use
JEWN BUG
An AI that reads procurement portals so a small painting crew does not have to.
Jewn Enterprises is a Black-owned painting company in West Roxbury, run by Nicole Wiggins. Public work is where a crew like hers grows, but the bids are scattered across city, state, and federal portals, each with its own format. Reading them all is unpaid, full-time labor that a working owner cannot do.
JEWN BUG is built around a single daily habit: open the app, see what is new, act on it. Overnight it searches the portals that matter to a painting contractor, parses each listing into plain fields, runs a verification pass so dead links and stale bids never reach the top, and ranks what is left by fit.
JEWN BUG is growing from finding the work to helping win attention for it. With email marketing, Nicole describes a campaign in plain English, the tool drafts a branded email, and she edits it, sends it to a list of her own contacts, and sees who opened it. Social media marketing is next.
Co-designed with Nicole, tested in her real routine, and reshaped by what she told us. A small, sturdy stack: a web front end on a lightweight API, with a search-grounded model doing the reading and a verification step checking its work.
One morning habit instead of unpaid full-time searching. It became one of the first real-world case studies for what grew into Project Lookout.
Good morning, Nicole.
3 painting bids worth your morning
-
- Painting scope, interior and exterior
- Municipal work inside the service area
- Link verified live this morning
-
- Interior repaint, a strong match
- Link verified live this morning
- Summer-only work window. Check the crew calendar.
-
- Includes a painting sub-scope
- Missing: DBE certification. Shown on purpose, so the gap is a to-do and not a mystery.
Verified links · 16 portals · refreshed overnight
Under the hood
A Python API and a React app in one container. A search-grounded model does the reading, structured government data fills in what it misses, and a verification pass checks the work.
- Frontend
- React 18ViteTailwind CSS
- Backend
- Python 3.11FastAPISQLAlchemy
- AI
- Gemini + Google Search groundingClaude fallbackJSON-structured output
- Data
- Firebase FirestoreSQLite cacheSAM.gov bulk data + API
- Transactional email serviceDaily digestsCampaignsOpen and click tracking
- Integrations
- CRM syncAccounting syncCloud document exportOAuth sign-in
- Infrastructure
- Google Cloud RunDockerCloud Scheduler
- Quality
- pytest245 automated tests
Tool · Live · Processing real purchase orders
Unicorn PO Dispenser
Purchase orders in a dozen formats, turned into clean records without anyone retyping a thing.
UNICORN is a Massachusetts women-owned distributor led by Ellen Cynar. Growth brought a problem that got worse the faster the business moved: purchase orders arriving from every customer in a different shape, as PDFs, as forwarded emails, as portal exports, each one demanding manual re-entry into the CRM, the accounting system, and a shared billing inbox.
The Dispenser takes the order in whatever form it arrives, extracts the fields, validates them against what the business already knows, flags anything that does not add up for a person to decide, and files a clean record where it belongs. A human still approves anything unusual, which is the point.
I co-designed the workflow with Ellen and her team around how they actually work, not around a vendor's template, then built it and shipped it. It runs in production and scales alongside the business it was built to keep up with.
The Hope Group didn't show up and lecture us. They listened first, then built something around how we actually work.
- PO number
- 44817
- Customer
- Aunty Stell
- Qty
- 24 cases
- Total
- $1,932.00
Flagged for a person.
- → CRM
- → Books
- → Billing
Filed. Nobody retyped a thing.
Under the hood
A React app on a small Node server. The model only extracts; the server holds every key and talks to the business systems, and a person approves before anything is filed.
- Frontend
- React 18ViteRecharts
- Backend
- Node 20ExpressServer-side API proxies
- AI
- Google GeminiStructured JSON extractionPDF, email, and text intake
- Auth + data
- Firebase AuthCloud FirestoreGoogle sign-in
- Integrations
- CRMAccounting systemEmail inboxCarrier shipping webhooksOAuth 2.0
- Infrastructure
- Google Cloud RunDocker
Open source · Live at cmd-z.com · Running the Lookout beta
Software drifts. Driftwood catches it.
An open source design research platform for running a beta without losing the plot.
We were running a beta out of a spreadsheet, and it was not going well. The bug tracker was a sheet everyone ignored and a chat channel that swallowed every report by Tuesday. We wanted one page a non-technical tester could open, point at what is wrong, and get on with their day, and one board the team could actually run the fix from.
That is all Driftwood is. A calm Stream for raw reports, a board with real working lanes, and two drawers for the wild ideas that always show up and never fit. It runs from one container and one config file, and it runs the live Project Lookout beta today.
Designed and developed end to end. Python, FastAPI, React, and Postgres in one Docker image on Cloud Run. Self-hosted, with no seat licenses.
The Stream 3 adrift
The morning digest shows yesterday's date.
Digest · screenshot attached
Steer it toTidy the column labels on the saved-leads export.
Pipeline · from a tester
Steer it toWhat if winning a contract set off confetti?
Idea · from a tester
Steer it to
The Stream is clear. Nothing adrift.
Bugs 0
Chores 0
In Progress 0
Completed 0
Three reports adrift. A person sights each one and steers it. Nothing is triaged automatically.
- 32API endpoints
- 22React views and components
- 14Hand-built animations, no UI kit, all reduced-motion safe
- 1Docker image, one command to deploy
Under the hood
One Docker image holds the API and the built app. Every row carries a tenant, so one board is real and each demo sandbox is its own throwaway copy.
- Frontend
- React 18ViteTailwind CSSReact Routerdnd kit
- Backend
- Python 3.11FastAPISQLAlchemy 2Pydantic
- Data
- PostgreSQL on Cloud SQLSQLite in developmentTenant-scoped rows
- Auth
- bcryptJWT in httpOnly cookiesRole-based access
- Integrations
- Token-gated feedback ingestRead-only Lookout analytics views
- Infrastructure
- Google Cloud RunMulti-stage Docker buildSecret Manager
Project Lookout · Patent pending · Live since October 6
Project Lookout is live, and BECMA is its first network.
Everything above led here. Project Lookout is the procurement-intelligence platform I designed and developed at the lab. It sweeps more than a hundred grant, contract, and bid sources every night, verifies every link, and scores each opportunity against a member's real business profile with a formula it can explain.
We built it to be run by networks, not sold seat by seat. The first is the Black Economic Council of Massachusetts, which now runs its own instance for its members. It is called BECMA Compass.
The day, in motion Mass Black Expo · October 6, 2026
Project Lookout is live, and BECMA is its first network
On October 6, the Black Economic Council of Massachusetts filled the Omni Boston Hotel at the Seaport for the 2026 Mass Black Expo. General admission sold out four days early. Fifty Black-owned businesses from across the Commonwealth ran the pavilion floor.
And on the main stage, Rev. Christopher Hope and I got to say something out loud that we had been building toward for a long time. Project Lookout is live, and it has its first network.
Chris made that case. Then I opened BECMA Compass on the big screen and ran it live, down to asking the assistant a real question: which of my saved leads is due soonest?
We opened with two numbers
From the city's own disparity study, as The Boston Globe reported in February 2021.
You cannot win what you never saw.
Why the first network matters
BECMA members helped shape the product. It was designed with them, tested inside their businesses, and changed by what they told us. So the name on the screen that morning was BECMA's, with “powered by Project Lookout” in the small print underneath. That is the right order.
The first month
One month in beta, with the first ten member businesses.
- 29,973opportunity matches put in front of members, each scored against a real business profile
- 2,503distinct grants, contracts, and bids behind those matches
- $9.2Bin stated value across the opportunities surfaced
- $14.4Min stated contract value in members' open pipelines, saved or bid on
- 2,517matches that carried a supplier-diversity set-aside
- 265cities, towns, and places reached
How to read these: platform-tracked figures from a beta, plus outcomes members reported themselves. Dollar values are what the postings state, not awards. The number we care about most is contracts won, and public procurement moves slower than a month.
Under the hood
A relay between code and AI. Models search and read; code verifies every link and computes every score. Multi-tenant from day one, so a new network comes on without a rebuild.
- Frontend
- React 18ViteTailwind CSSLight and dark themesEnglish + Spanish
- Backend
- Python 3.11FastAPISQLAlchemy 2Alembic migrations
- AI
- Gemini grounded searchFact extractionProposal draftingDeterministic scoring in code
- Data
- PostgreSQL on Cloud SQLMulti-tenantSAM.gov, Grants.gov, COMMBUYS feeds
- Auth
- Invite-only accountsGoogle + Microsoft OAuthJWT in httpOnly cookiesbcrypt
- Infrastructure
- Google Cloud RunCloud Run JobsCloud SchedulerSecret ManagerSendGridCloud Monitoring
- Quality
- 1,400+ automated testspytest + VitestAccuracy and bias gate on every release
Designed and developed by Stephen Walter Produced by Rev. Christopher Hope Quality tested by Yunzi Wan Inspired by a collaboration with Nicole Wiggins of Jewn Enterprises
The Hope Group · An ethical AI lab in Boston · hopegroup.ai
