AI proposes. People decide.
AI research and CRM workflows for non-technical teams. Confidence-based review and explicit confirmation before any Salesforce write; 88% of AI-proposed changes accepted as written.
60-second summary
- Who:
- Non-technical outreach and advancement teams at a photography talent agency and two schools.
- Key decisions:
- Models read and categorize, rules match; staff set the auto-approval threshold; staff confirm every Salesforce write; enrichment never changes consent.
- Results:
- 88% of proposed changes accepted as written, 12% edited or rejected; 3,000+ missing emails recovered; agency-observed bounce rate down from 2.6% to 0.6%; a 160-person pilot became a paid 5,078-record engagement.
- My role:
- Founded it and built it myself, part-time alongside Swish; collaborators handled QA, delivery coordination and commercial work.
I founded and personally built Agency Edge: Python/Flask services, SQLAlchemy models, Celery/Redis workers, React/TypeScript review UI, and LLM steps on the OpenAI API with structured outputs and function calling. Collaborators handled QA, delivery coordination and commercial work.
Shared research and review capabilities serve three paying organizations: a photography talent agency, Marin Academy and San Francisco University High School. Customer-specific matching rules stay separate.
88% / 12% AI-proposed changes accepted as written / edited or rejected in review
2.6% → 0.6% email bounce rate for the agency’s campaigns, 2024 to 2025
3,000+ missing email addresses recovered for existing agency contacts
160 → 5,078 records, from Marin Academy pilot to paid engagement
Research, review, confirm
One workflow that keeps people in control of their own records.
Package research as reusable steps
Teams run research as named, reusable steps, such as refreshing a contact’s profile, finding a verified email, or gathering company context. Each step can run on demand or automatically for new and stale records, and every run is saved with its sources and the changes it proposed.

Route each change by confidence
Every proposed field change carries a confidence label. A staff-set confidence threshold decides whether a change is auto-approved inside Agency Edge; everything below it goes to a review queue. Staff compare proposed and current values, edit if needed, and approve or reject.

Quality loop
When Marin Academy staff found wrong-person matches in the pilot, we audited and corrected the records, added confidence information so staff could see which matches needed review, and split interests into categories they could filter. The review queue now records whether each proposed change is accepted as written, edited or rejected: 88% are accepted as written and 12% are edited or rejected.
Confirm before anything reaches Salesforce
Before an update is written to Salesforce, staff see Agency Edge’s values side by side with what’s in Salesforce now, then confirm the sync. Approving a change inside Agency Edge, confirming the Salesforce update, queued work and results are separate states, so staff always know what has actually changed in their system of record.

Photography talent agency
Photographer prospecting in the tools staff already use.
The agency needed to connect its photographers' work with relevant people across about 40,000 CRM contacts. Staff needed to understand a prospect’s company and work, see which photographers fit, and decide how to follow up.
Decide where AI belongs
Search and website crawling gathered contact and company context, and models extracted fields and assigned business categories. Matching stayed in plain application rules: the agency supplied its own category taxonomy and thresholds, and I encoded rules combining category overlap, role fit and each photographer’s campaign engagement. That kept every match explainable, instead of asking a model to rank the whole CRM.
Explain every match
Each photographer tag carries its reasons: shared categories, role fit, campaign engagement and links to the work behind it. Staff can check why a tag appears before deciding how to follow up, and agency staff reported using these explanations in prospect follow-up. Engagement and fit are scored separately, so staff can see how a tag was assigned.

Fit the tools the team already uses
Research results went back into Salesforce and Campaign Monitor as enriched fields and photographer tags, so staff could build segments without moving data between dashboards. The first version wrote these updates automatically; I later added review and an explicit confirmation step before any Salesforce write.
Respect consent
Enrichment filled in a contact’s title, LinkedIn and location, but never changed whether they could be emailed: an existing opt-out stayed in place.
Results the agency observed
- 2.6% → 0.6%Bounce rate
- 3,000+Missing emails recovered
- 6,000+Employer records updated
Bounce rate as observed by the agency, 2024 vs. 2025 campaigns.
The contact work recovered missing email addresses for over 3,000 existing contacts and updated employer information for over 6,000; these groups may overlap.
More agency product detail

Marin Academy
Turning pilot feedback into a paid engagement.
Marin Academy staff wanted alumni context they could use for engagement preparation. I built and delivered a 160-person research pilot. Staff feedback surfaced wrong-person matches and a need for interests they could filter, instead of information buried in a biography.
- 160-personresearch pilot
- Staff feedbackchanged the work
- Paid 5,078-recordresearch engagement
| Staff feedback or need | Product and delivery response |
|---|---|
| A possible wrong-person match | Audit and correction of pilot records, with confidence information to support review. |
| Interests useful for outreach and student mentorship | Separately usable interest categories, instead of relying only on narrative biographies. |
| Results that match school records | Preserve the school’s identifiers and original/researched comparison context. |
The paid 5,078-record engagement preserved the school’s identifiers for review and matching. I owned the software and product changes; collaborators contributed comparison formatting, audit, manual QA and delivery coordination.
San Francisco University High School
Research delivered in the format staff review in.
San Francisco University High School (UHS) staff asked during the pilot to compare original and researched information in one view. We added a combined comparison tab while keeping the original views. Stable constituent IDs provided a matching key for the school’s Raiser’s Edge records.
- Constituent ID
- Original information
- Researched information
The school owned review and any import into its CRM.
How to read these numbers
- Scores and confidence labels are signals inside the application, not measured accuracy.
- Customer results are what customers observed during the work, not controlled attribution studies.
- Candidate emails are research findings, not verified deliverability or a count of people reached.
- Screenshots come from different customers and product versions; they are not one continuous transaction.
Let’s talk.
I’m looking for product roles building AI for knowledge workers, working closely with customers, engineers and research teams to turn complex workflows into products people trust.

