Built for Data for Good Calgary

A whole volunteer operation on one Oracle platform

Intake, two CRMs, projects, e-signature, meetings, a public website and private AI, designed, built and run by DB Realm on Oracle APEX 26.1 and Oracle AI Database 26ai on OCI.

Member record
AC
Avery Chen
avery.chen@example.org
Data visualizationSQLPython
Project: Riverbank Cleanup Dashboard
  • Expression of interest Received
  • CRM profile Matched
  • NDA Signed
  • Workspace account Active
  • Attendance 5 of 6 meetings
Mock-ups. People, projects and values are fictional.
pages across two Oracle APEX apps
185
pages across two Oracle APEX apps
modules sharing one data model
11
modules sharing one data model
PL/SQL packages written for the system
24
PL/SQL packages written for the system
records searchable by meaning
2,295
records searchable by meaning

What the platform delivers

One source of truth

CRM records, forms, signatures and projects meet in one database instead of six separate tools.

Private, accountable AI

An on-premises model answers questions from the data, suggests matches and logs every call.

Consent by design

Do-not-email contacts are always excluded, and live sending is off until an admin enables it.

Run like production

A web firewall, a full audit trail, a monitor for silent failures and tests that gate every release.

Built with
  • Oracle APEX 26.1
  • Oracle AI Database 26ai
  • Oracle Cloud Infrastructure
  • Oracle AI Vector Search
  • PL/SQL
  • APEXlang in git
  • utPLSQL + Playwright
How it works

One database at the centre of the operation

Every service connects to the database; people reach two apps through one firewall.

Connected services

  • HubSpot CRM
    Contacts and companies, every 6 hours
  • Mautic
    Volunteer CRM, mirrored and pushed
  • Google Forms
    Expressions of interest
  • DocuSeal
    NDA e-signature and status
  • NextCloud
    Project accounts and groups
  • LLM server
    Private, on-premises

On Oracle Cloud

  • Oracle APEX 26.1
    Staff app and public site
  • AI Database 26ai
    Autonomous, private endpoint
  • PL/SQL packages
    24 written for the system
  • Scheduler
    12 jobs, monitored hourly
  • AI Vector Search
    2,295 embeddings, HNSW
  • Load balancer + WAF
    The only way in

People

  • Core team
    Projects, members, meetings
  • Project leads
    Rosters, attendance, NDAs
  • Administrators
    Access, settings, audit
  • Volunteers
    Apply, join, sign NDAs
  • Partners
    Suggest projects, portal link
  • Public visitors
    Resources, open positions

Integrations run as PL/SQL packages inside the database; their credentials are stored encrypted. Third-party names identify services the platform connects to; DB Realm is not affiliated with their owners.

Platform modules

Eleven modules, one data model

Every module reads and writes the same Oracle schema, so a volunteer, a project or a signature means the same thing on every page.

Project management

Partner and micro projects, members, roles, NDA status, team accounts.

Intake & positions

Expressions of interest, dated recruitment rounds, AI screening.

Resource library

Curated catalog, moderation queue, click tracking, link checks.

HubSpot mirror

Contacts, companies and engagements, pulled every 6 hours and reconciled.

Mautic mirror

Volunteer CRM with tags, do-not-contact, a map and semantic search.

Micro-projects

Ready to launch

Suggest, approve, advertise, recruit, assign, deliver, showcase.

Email & campaigns

Send gated

Templates, brands and consent-floor campaigns.

AI assistant

Private model, grounded answers, vector search and a one-click model switch.

Security & admin

Own sign-in, email codes, feature-level access, audit trail.

Monitoring

Activity, errors, email and AI logs, link health and backend job alerts.

Public site

Volunteer forms, NDA, collaboration requests, resources, feedback.

On the roadmap
77

features tracked on a roadmap generated from one data file

tables in one schema
112
tables in one schema
PL/SQL packages written for DfG
24
PL/SQL packages written for DfG
scheduled jobs, monitored hourly
12
scheduled jobs, monitored hourly
authorization schemes in the staff app
18
authorization schemes in the staff app
The volunteer path

From first form to finished project

Eight steps on the same volunteer record. Each step is a module of the platform, and each hand-off is automatic or one click.

  1. 01 Forms · public site

    Expression of interest

    A volunteer applies through a form or the public site; the answers land in the database.

  2. 02 HubSpot · Mautic

    CRM mirror

    HubSpot and Mautic profiles are mirrored in, so skills, tags and consent are on hand.

  3. 03 AI Vector Search

    AI-assisted matching

    Vector search ranks volunteers against a project's needs. A person decides.

  4. 04 Projects

    Project assignment

    The core team approves and assigns; the volunteer joins the project as an invited member.

  5. 05 DocuSeal

    NDA e-signature

    Projects that need an NDA send it for signing; the status syncs back every night.

  6. 06 NextCloud

    Collaboration accounts

    Workspace accounts and project groups are created for the team.

  7. 07 Meetings

    Meetings & attendance

    Meetings are scheduled, rosters generated, attendance recorded and reported.

  8. 08 Ready to launch

    Showcase

    Finished projects are published with their outcome and an opt-in team credit.

Project showcase
Riverbank Cleanup Dashboard

A volunteer team turned cleanup logs into a public map of litter hot spots for a riverside society.

Estimated
4 weeks
Delivered
5 weeks
Team (opted in): Avery C., Jordan P., Sam R.
Mock-ups. People, projects and values are fictional.
AI built in

Private AI that cites its data and logs every call

Five AI features help staff ask questions, screen applicants and match volunteers. All of them run on a private model, read the data they cite, and leave a record.

Private model

Models run on a private on-premises server. Volunteer data never goes to a public AI service.

Search by meaning

Oracle AI Vector Search: 2,295 embeddings across 9 record types, HNSW index.

Read-only answers

A question becomes one read-only query over approved views, checked again before it runs.

Advisory only

Screening and matching suggest; staff decide. AI never changes a record on its own.

Calibrated scores

A match shows its rank and the pool's typical score, never a bare percentage.

Every call logged

Prompt, response, feature and duration of every AI call go to an admin-only log.

Ask the data
Which active projects still need someone for data visualization?
Two active projects list data visualization as an open need: Riverbank Cleanup Dashboard and Library Usage Trends.
Grounded: 1 read-only query Logged
Suggested volunteers · advisory
Library Usage Trends · needs SQL, data visualization
  • AC
    Avery Chen Best of 240
    61% · pool typical 54%
    Basis: CRM profile
  • JP
    Jordan Patel 2nd of 240
    59% · pool typical 54%
    Basis: CRM profile
  • SR
    Sam Rivera 3rd of 240
    57% · pool typical 54%
    Basis: declared skills
A person approves every assignment.
Mock-ups. People, projects and values are fictional.

One switch, every feature: an admin picks the language model for all five AI features from one list. Today: gpt-oss-120b for answers, qwen3-embedding for search.

Public engagement

Open to the public, careful with consent

The public site brings volunteers, partners and visitors in. The email tools only ever reach people who said yes.

Resource library

A searchable catalog of data tools and learning, with facets and broken-link flags. Visitors can suggest new resources.

Open positions

Recruitment rounds open and close on their dates; the Positions menu only appears while a round is open.

Micro-projects

Ready to launch

Organizations suggest a small data project; staff review and advertise it; volunteers join. Owners use a private link.

Email & campaigns

15 branded templates with live preview and versions. Campaigns preview every recipient's email before anything is sent.

Two separate choices

Sharing contact details and joining the mailing list are separate, unticked boxes.

A consent floor, and a gate

Do-not-email and unconsented contacts are removed in the database, every time. Live sending is off until an admin enables it; tests go to an admin inbox.

Join a micro-project
Library Usage Trends
First name
Email
Share my contact details with the project owner. If not, Data for Good makes the introduction.
Add me to the volunteer mailing list.
Register interest
Campaign · Autumn volunteer update
City: Calgary Skill: Python
  • Recipients (consented) 128
  • Excluded: do-not-email 7
  • Excluded: no consent on file 12
  • Previews built 128 of 128
Live sending: off (dry run)
Send test to me Send campaign
Mock-ups. People, projects and values are fictional.
Security & reliability

Run with production discipline

Least-privilege access, a full audit trail, and monitoring that catches silent failures.

Access by feature and role

18 authorization schemes; every page, button and link is gated to match.

Passwordless sign-in

Email codes: 6 digits, 10 minutes, 5 tries. Only their hashes are stored.

Audit trail

Admin actions and sign-ins in one trail; access links revocable from a monitor.

Web application firewall

OCI WAF in front of both apps, with OWASP rules, rate limits and bot blocking.

Backend job monitor

Hourly checks of 12 jobs for loud and silent failures; one alert per problem.

Link health

Every resource link checked weekly, with run history and one-click re-check.

Why the job monitor exists: a scheduled sync once reported success on 100 runs in a row while pulling nothing for 25 days. The monitor now checks what each job pulled, as well as whether it ran.

Quality

Four test layers before every release

The full gate runs on a copy of the system before every release, so production only ever receives a change that has already passed.

  1. 4

    Browser tests

    170 Playwright tests on the twin; 39 read-only on production.

  2. 3

    Region sweep

    109 pages of both apps rendered and checked for any error.

  3. 2

    Database unit tests

    64 utPLSQL tests on the PL/SQL the pages rely on.

  4. 1

    Fast gates

    5 source checks on every commit.

A non-production twin

Tests that write data run only on the twin. Three independent controls refuse production.

How we deliver

The habits we bring to every engagement

The practices that keep this platform healthy are the ones DB Realm brings to every system it builds and runs.

  1. 01

    Source in git, not only in a browser

    Every page of both apps lives in git as APEXlang text, so changes are reviewed, compared and rebuilt like any other code.

  2. 02

    Independent, versioned releases

    Each app carries its own version and only the app that changed is imported, so staff-app work never disturbs the public site.

  3. 03

    A backup on each side of every release

    The app is exported before and after each import, so every version has a restore point.

  4. 04

    A roadmap generated from data

    77 features across 11 modules live in one data file; the roadmap page is rebuilt from it and checked on every commit.

  5. 05

    Agent-assisted, verification-gated

    AI coding agents do much of the building. Gates, tests and live checks decide what ships, and each lesson learned is written down.

  1. Dec 2025
    First commit
  2. Jul 2026
    Security and administration rebuild
  3. Aug 2026
    Backend job monitoring
  4. Oct 2026
    Release gates on every change
Who built it
Fekrat El Wehedi
Director of Technical Operations, Data for Good Calgary · Founder, DB Realm

The platform was designed, built and is run by the chapter's own Director of Technical Operations, through DB Realm. One person owns all of it: the architecture and data model, every page of both apps, the integrations, the private AI, the cloud infrastructure, and the release gates that keep it safe to change. It was built from inside the organisation, around how the chapter actually works.

The full story

Data for Good Calgary — Volunteer Platform

Data for Good Calgary

Data for Good Calgary matches volunteer data scientists, analysts and engineers with non-profits that need help making sense of their data. Demand was never the chapter’s constraint. Coordination was: running a volunteer organisation across a stack of disconnected tools.

Two applications, one database

The platform runs on Oracle AI Database 26ai with Oracle APEX 26.1 in front of it: a staff application for the core team, project leads and administrators, and a public site that volunteers, partners and visitors use without an account. Both read and write one schema of 112 tables, with 24 PL/SQL packages written for the system.

All application source is kept in APEXlang (.apx) rather than binary exports, so pages are reviewable in a diff and deployable through SQLcl like any other code.

Automating the onboarding chain

The work that used to be manual is now driven from PL/SQL:

For NextCloud, the database calls the OCS API to create accounts and keep project group membership in step as volunteers join and leave projects. Account passwords are generated with DBMS_CRYPTO.RANDOMBYTES rather than a pseudo-random source, because those accounts give access to partner data.

DocuSeal is driven over REST, issuing NDAs from managed templates and pulling submission status back every night. The signed-field JSON is unpivoted into queryable views, so compliance reporting is a SQL question rather than a folder audit.

HubSpot is pulled every six hours and each run stores HubSpot’s own total beside the local count; Mautic is mirrored nightly and new volunteers are pushed back. A discrepancy surfaces the next morning instead of at the next audit.

Engineering details that mattered

The NextCloud OCS API returns two status codes: the HTTP status, and an OCS status inside the response JSON. An HTTP 200 with an OCS 102 means “user already exists,” not success. Every call checks both. An integration that checks only the HTTP status quietly does nothing.

The same lesson shaped the job monitor. A scheduler job can report success without doing its work, so the monitor checks results (rows pulled, totals reconciled) as well as exit status, and raises one alert per problem rather than one per run.

Live

This platform is one of three connected pieces of work for the chapter. See also the marketing automation migration and the cloud infrastructure underneath it.

Ready to Build Something Similar?

Let's discuss how we can bring the same expertise to your organization.

Start a Conversation