source: .forge/discovery/meeting-notes/2026-10-16-kavindu-progress-update.md · generated — edit the markdown, not this file

Empire Teas CRM — progress update with Empire Teas

Date: 16 October 2026 Meeting: online call, about sixteen minutes — the progress meeting Rukshan asked for (2026-10-14-kavindu-search-provider.md §3) Participants: Kavindu Anthony (Empire Teas); Tharinda Rajapaksha, Dinil Gamage, Nanndish Satgunarajah and Thilina Abeyrathna (Altrium) Source: the meeting tool’s transcript and summary, read on 16 October 2026. Where the two disagree, this note follows the transcript: the summary gives Tharinda the action of speaking to Sajeev De Silva, but in the transcript Kavindu says he will speak to Sajeev himself.

Why this file exists: it is the dated record of Empire’s first look at the CRM on dev, of the questions Kavindu asked about the AI layer and the mailbox sync, and of the follow-ups agreed. Three things said on Altrium’s side do not match the design or the issued proposal; §6 lists them, and .forge/owed.md carries the correction.


1. The web-search tool

  • Kavindu had read the search-tool comparison and asked about its 42% accuracy figure. Dinil explained that the test set mixed well-known companies with little-known ones, and some companies have almost no online presence, so every tool missed some; the aim of the test was to see where the tools differ. (The comparison is .forge/design/ai/lead-research-source-options.md §11.)
  • Kavindu asked what accuracy to expect in production, for example on cards scanned at a trade fair. Dinil said no figure can be given yet: it depends on each company’s online presence, and the test used Empire’s existing customers.
  • Kavindu’s definition of accuracy: when the tool finds nothing, the CRM must say it found nothing, and must never show false information. Dinil confirmed that the CRM will tell the user no reliable results were found, and will not rank a lead lower because nothing was found about it. This is already the design: lead-research-source-options.md §11 (findings, item 4) makes “nothing found” a normal outcome with its own state, and the F11-g stub scores such a lead from the inquiry alone at lower confidence, not lower priority.
  • Kavindu will speak to Sajeev De Silva about the search tool and answer on Exa. The provider choice still waits on that answer (.forge/owed.md, the search-provider row).

2. The AI layer

  • Kavindu asked where else AI is used. Dinil and Tharinda named lead research and lead ranking (F11-c, F11-g) and email drafting (F11-h) for Phase 1.
  • Tharinda said the models are Claude on Amazon Bedrock. Thilina said the team uses two models: Haiku for ranking, drafting and other routine work, and Sonnet for complex work, and that the newest Haiku, released the week before, is accurate and fast. (As of 16 October the AWS account can call Haiku 4.5 only; Haiku 5.5 and Sonnet 5.5 are refused until AWS raises their quotas — .forge/owed.md, the Bedrock row, and cloud-gated-validation.md §C-5.)
  • Kavindu asked whether the models keep learning from the data, and whether the data stays inside Empire’s environment. Tharinda answered yes to the first; Thilina said no identifiable data is sent to the model, only what ranking and drafting need, and that the prompts are not written yet. Both answers differ from the design and the issued proposal (§6).

3. Progress and Phase 1 scope

  • Tharinda summarised the work so far: Microsoft sign-in, the first part of the email integration, and the CRM deployed on dev at https://dev.crm.empire.lk. The UI still has issues, and the AI features are waiting on open items.
  • Phase 1 target: end of November. Its main features, as stated: users and roles, accounts, leads, email, and the weekly prices. (The build is due 20 November and the hand-over is Fri 27 November — .forge/client/project-timeline.md.)
  • Kavindu cannot sign in to dev with his own address; Empire’s test mailboxes (CRM01 to CRM05) are the sign-in accounts there. This is as designed: sign-in needs a user created in advance (AD #34).

4. Demonstration

Nanndish signed in as CRM03 and showed, on dev:

  • the seeded IT Admin role and the first user, and a role created on the Roles screen with every permission, given to CRM03;
  • creating an account;
  • creating a lead by hand, and the New lead from inquiry screen, where an inquiry is pasted and extracted. Extraction is not built yet (F12);
  • the email integration (Outlook);
  • the pricing screen, where standards are added;
  • the Users and Roles screens.

5. Mailbox sync

  • Kavindu asked whether the sync brings in every email or only the relevant ones. Dinil answered that it brings in every email.
  • Kavindu asked whether keeping unrelated mail is worthwhile and whether it uses space, since some mailboxes are over 50 GB. Tharinda said the 50 GB would not be copied and the sync should bring in only the mail from the user’s leads, and that the team will look into a filter. Kavindu agreed there is no point in loading every email.
  • What F20-b does today (.forge/specs/F20-b-inbound-sync-inbox-spec.md): it takes in all mail in both directions except Drafts, Junk Email and Deleted Items (FR-3), back-fills the 30 days before the connection, a setting (D-2, FR-5), and then every new message. Each message is stored in the CRM database, body included (FR-2, FR-4). So a mailbox’s history is not copied, but every message since the connection is kept, whether or not it concerns a customer. F20-f, still a stub, links mail to the matching contact, account and lead and leaves unmatched mail in the inbox.
  • Action: Tharinda looks into limiting the sync to lead-related mail (.forge/owed.md).

6. Statements to correct with Empire

Three answers given in the meeting do not match the design or the issued proposal. The proposal outranks internal documents and in-meeting answers (.claude/rules/source-trust.md), so the record is corrected with Empire rather than left to stand. .forge/owed.md carries the row.

What was said What the design and the proposal say
The models keep learning from the data (Tharinda: “Yes”) They do not. On Bedrock the models run in accounts AWS owns, and Anthropic has no access to them, so it never sees Empire’s prompts or outputs (AWS, Amazon Bedrock User Guide, “Data protection”, read 2026-10-16). AWS also states that neither AWS nor the model providers use Bedrock inputs or outputs to train models (Amazon Bedrock FAQs, security section, read 2026-10-16). The models are pinned versions (AD #19) and do not change from use. (Proposal §A.6.4 says the same of Anthropic’s own API terms, written before Bedrock was chosen.)
No identifiable data is sent to the model (Thilina) The task’s own content is sent. Proposal §A.6.1: the one email being processed (extraction), the lead’s own fields (triage), one account’s activity (briefing), the thread and account context (drafting). Every call is made from Empire’s own AWS account (AD #19). Where Sonnet 5.5’s requests are processed, possibly outside the US, is OQ-15, still open with Empire (client question Q18). Settled 4 Nov 2026: any commercial AWS Region is accepted; Q18 answered and closed (PRD Rev 82). If Exa is chosen, the company searches also leave AWS (lead-research-source-options.md §12).
The synced mail uses no space in the database (Tharinda) Each synced message is stored, body included (§5).

7. Detailed presentation

Kavindu asked for a detailed presentation when Empire’s sales team is back, with Praveen and others. The team returns around 24 October; the session is to be in the last week of October (26 to 30 October).

Actions

Action Owner Recorded in
Speak to Sajeev De Silva about the search tool, and answer on Exa Kavindu Anthony .forge/owed.md (the search-provider row)
Update Sajeev De Silva on this meeting Kavindu Anthony this note
Look into limiting the mailbox sync to lead-related mail Tharinda .forge/owed.md (the mailbox-sync row)
Arrange the detailed presentation for the last week of October, with Empire’s sales team Tharinda .forge/owed.md (the presentation row)
Correct the three statements in §6 with Kavindu Tharinda .forge/owed.md (the AI data-handling row)

Generated from .forge/discovery/meeting-notes/2026-10-16-kavindu-progress-update.md by .claude/lib/render-docs-html.js. Markdown is canonical.