redmine/docs/contact_triage_loop_raw_notes_2026-07-04.md

eh, I think you misunderstood me about the filters. I see this:

batch-update Apply one payload to a filtered contact batch.

Unless I'm misunderstanding how to use that. it sounds like it's not what I wanted/

I was hoping to be able to do this:

1. In batches of perhaps 10-20, pull a list of contacts that have not been updated for a while, and who either:

   • also have specific tags
   • also lack specific tags

2. For each of those contacts, do the following:

   • check if there's anything about the contact itself that informs me I shouldn't need to look at anything else about it, and I can quickly tag them as something like "spam". I expect most of the time, I won't be able to do that.

   • Fetch the 3-5 most recent tickets from the contact, and look at them, using some framework to score any tags that should apply

     • Examples:

       • customer:

         • somebody on our side emailed back +1
         • seems like they're talking about packages or items in our warehouse +1

       • promo + spam: contact seems like a business sending us advertisements

         • from address is something like "no-reply", "noreply", "deals", or "promotions" +1
         • it mentions a deal, discount, or coupon +1

   • If the scores for any particular set set of tags >= a threshhold, use add_tag to add those tags to the contact

     • otherwise, add a tag like "ai-checked"

   • The last step caused the updated date of the contact to increase

3. Due to the very last step above (which bumps the updated date), if I wait and loop back through Step 1 again a few minutes or hours later, I won't retrieve the same contacts I already tagged.

   • In this way, a looping/repeating task will eventually cycle through all contacts.
     • Eventually, the ones it has already tagged will have been updated long enough ago, that they will get pulled again

To be clear:

I am describing a broad pattern of workflow here, not specifics. I am just explaining this because:

a) The way you implemented the changes seems counterintuitive for how I want to use this, but I could of course be misunderstanding.

b) I think the basic concept of this loop should be added to the project documentation, maybe even with a diagram

Further down the road:

• I will want to develop tags that indicates a specific kind of evaluation needs performed with a contact, just as summarizing their most recent communications with us.
  • When looping through contacts, I can add that tag to contacts whose comms need to be summarized
  • Then, another agent or script can fetch only contacts tagged to indicate they need summarized
    • That script/agent fetches their recent communications, does it's job, and removes the tag, indicating that job is finished
• Other likely tags I will eventually implement:
  • basic info needs corrected:
    • Typical auto-import problems exist in the contact data, such as:
      • company field contains their email address
      • username portion of email address is in first_name, but their actual name is in the footer of the email or in the "from" part
      • basic contact info is actually in the ticket, but not in the correct contact fields:
        • first name, last name
        • additional email addresses
        • phone number(s)
        • job title
        • name of their business
      • We know the name of the company they work for, but no contact.is_company=true exists with that name
        • Create it
        • Tag it as needing enrichment
        • Update the contact(s) who work for that company, setting company = that company
  • NMI mapping:
    • Contact seems to be a customer, but their "Account Number" field is blank, so we're not sure which NMI account is their's
      • Try to figure it out by comparing stuff like:
        • email address
        • name of company
        • tracking numbers or order ids mentioned
        • ^^ above, but for other contacts who work at the same company
      • Update linkages all around:
        • Update that company (and it's employees) contacts in redmine so "Account Number" matches user_account.user_account_id in NMI
        • Set the redmine.contact_id's into a mapping table in NMI that link which contacts are for which account
        • Update obsidian profile about the contact/person to include both pieces of info
  • entity extraction and enrichment:
    • members of our team responded to somebody, but we don't have (much) info about them in either redmine contacts or in obsidian
    • This tells us the person is probably an active employee, so we:
      • search for 5-10 other tickets they worked on lately, and use that to infer a profile of what their job is or fill in context about what they've worked on lately
    • phone number linkage to accounts:

      • transcribe/ingest tickets that are call transcripts, adding summary to contact file automatically

        • add "note/call" to contact automatically
        • identify if follow-up action is needed; if it is, create ticket, assigned to relevant person

      • bobcat caller id: recognized contact calls us...

        • route call based on context, not just the number:
          • we're super busy, short staffed, and person likes to talk alot: "sorry, we're not available"
          • Bob usually talks to this person, but Bob is marked "do not disturb": "voice mail / not available"
        • show on-screen bobcat overlay caller-id
  • Needing enrichment (likely more than one tag to represent the kind of enrichment):
    • social media:
      • try to identify their social media profiles (linkedin, facebook, twitter, etc) using name, email, etc.
      • possible search known social media groups for people with same/similar name
      • identify product website(s)=
    • potential suppliers:
      • search places like alibaba for products we've seen them sell, and try to narrow down who their suppliers are
      • pull checkin pictures of shipping labels from them, and cross-reference with other info to locate possible suppliers
  • special handling:
    • SLA/VIP: this contact needs very fast or special responses for some reason
      • possibly tied to dedicated agent who does something special with their emails, like:
        • trace package: pull data we'd need to trace the workflows about it:
          • if it looks like it has a tracking number, kick off the workflow to pull that up in NMI, get pictures, check who worked on it, tracking status, etc.
          • Possibly fetch surveillance from synology autmatically:
            • For each major timestamp involved ( box.checkin_date, box.presort_date, individual_product.checkin_date ) pull 2-3 minutes of footage before/after from cameras known to overlook the 2-3 nearest locations in the warehouse
              • split the video feed into a series of still images 30-seconds apart
              • have a VLM explain what it sees in the images, even checking for specific things:
                • presence of truck at loading dock
                • person seen walking away from area carrying similar locking box/item
        • import order: email is an invoice, PO, or similar listing what a person is sending to us, perhaps the reason, and when it should arrive.
          • they're a vendor: log that supplies were being delivered and what we spent on them
          • prep customer: make a prep inbound order if it doesn't exist already
        • vendor receipt or invoice:
          • automatically log/input that into the expense tracker if it's not there already
          • check if there is a purchase order for it.
            • if there is, do something like mark it en-route, received, etc.
              • create ticket, assigning somebody to put it away
            • assign an AI-INFO-BLOCK to the manager, asking them to find out who ordered it, what for, and why they didn't make a PO
    • employee and it's about scheduling
      • check the attendance-schedule project to see if they already reported it in there
        • if not, add it
          • assign an AI-INFO-BLOCK asking them to explain why they could not do that themselves
        • check for time-off blockers: - [ ] somebody else already has time off within 1-2 days - [ ] customer has complained about delays in the past week - [ ] same person asked off on the same weekday/pattern in the past 3-4 weeks, and no good reason - [ ] overall billings/productivity have been low for over a week
          • if blocked: warn them they can't have the day off unless they can prove it's manditory/unavoidable
            • if they say it is, AI-INFO-BLOCK until they provide proof or admit they don't have it
  • very simple question
    • sales 20-questions: they sent an email that amounts to playing 20-questions about capacity (wanting to start project)
      • AI-agent uses obsidian and other data:
        • generate really slick looking brochure quickly, using obsidian-backed data
        • loop human in to approve it
        • update/expand based on human input, then send it over to client
        • save it for later, so next time, it can just be sent right away
  • spammer-be-gone:
    • Identify habitual or serious spammers, time wasters, and promoters emailing us
      • use gmail MCP to add filters to contact@ldrprep.com so that person's mail doesn't get imported into redmine anymore
  • high-risk contact:
    • there is some special risk or liability related to talking to that contact. For examples:
      • litigious person or debt collector
      • disgruntled former employee
      • stalker spouse of current employee
    • handle ticket(s) in a manner people who shouldn't talk to them can't see it or can't respond to it
      • phone calls from them do not ring to the general line anyone can answer
  • lead follow-up:
    • contacts marked as wanting to potentially buy something from us or start services, but we've heard crickets for 7+ days
      • encourage somebody to call/email them, maybe even suggest what to say/send

Please ensure this info is saved with the project and easy to locate

For right now, much of this will not be immediately practical to implement, but I think storing it here is both important context and saves time.