Baseer GTM — System Report

Three distinct content voices Corrected

• The content planner now writes in three clearly separate voices: Sara (bilingual Misraj founder), Rawan (Arabic business-development / digital-transformation), and Norah — corrected today to an independent Saudi AI enthusiast, not a Misraj voice.
• Norah's content was regenerated as broad AI interest — trends, tools, papers, and ideas she finds genuinely useful — with no product pitching and no Misraj/Baseer affiliation. This keeps her account authentic rather than a disguised brand channel.

Norah X auto-post Live · fail-closed

• Norah's approved, scheduled posts now publish to X automatically through twitterapi.io — no second browser, no manual copy-paste.
• Gated in order: off unless TWITTER_AUTOPOST_ENABLED is exactly on, a 5 posts/day cap, optional business-hours-only window, operator-approved items only, kill-switch/global-pause respected, and idempotent (never double-posts). Drafts never auto-post.

Signals: "reply as Norah" Live · approval-gated

• On any Twitter signal, the operator can now draft a reply in Norah's voice, edit it, and — after an explicit approval — have it posted as a real reply to that tweet.
• The send clears the same execution gate as every other outbound action (content safety, exact-approval match, reply-target guard, kill-switch, idempotency) and is off unless TWITTER_REPLY_ENABLED is on. Recorded in a dedicated table (migration 047).

Signals Hub: multi-source + recency Fixed

• The scan now pulls Twitter, news, and LinkedIn together — a bug that returned zero news was fixed by parsing the news RSS feeds server-side, so all three streams flow into one board.
• A new rolling recency rule (migration 048, default 7 days) drops stale content at every layer — scan, projection, and the surface — so the system only ever studies, replies to, or retweets recent content, never a months-old post it just happened to discover today.

Quality & data model

• ~1,458 automated tests passing (up from ~1,400), covering the Twitter auto-post and reply-as-Norah paths, the multi-source scan, and the recency gate.
• Database schema is now at migration 048 — the twitter-reply drafts table (047) and the signal recency column (048), both additive and fail-closed.

What's next

• Flip on Norah's auto-post and reply flags once the first approved items are queued, and watch the first live posts and replies land cleanly.
• Tune the 7-day recency window per source if a channel proves faster- or slower-moving than the default.

Contact-supply backbone Rebuilt

• The core gap is closed: an approved sourced account or a discovered contact now lands as a real net-new row (the sourcing → accounts → contacts projection). Before today, approved AI outputs were reviewed and then evaporated — nothing became a contact the outreach engine could act on.
• Contact-discovery is now scheduled daily and its runaway loops are fixed — tasks that used to churn to an 18-minute timeout now converge, with a per-task-type tool-round budget and a valid "found none" result.
• A broken linkedin_profile_activity tool (was passing the wrong identifier) was fixed, so the LinkedIn activity gate that disqualifies dormant profiles works again.

Email-first for 17 top executives Live

• An email-first exec gate now routes LinkedIn-inactive GCC decision-makers to email instead of a dead LinkedIn invite. A deterministic Apollo reveal-and-enroll pass ran on the stuck cohort.
• 11 of 17 executives had a verified email revealed via Apollo and were enrolled on the email channel (e.g. geidea, Bank Albilad, Etihad, Jeel, ALDAR).
• 6 had no Apollo email and were returned to the LinkedIn track — including a Deputy Minister who is now flagged for a manual, human-owned approach rather than automation.

Message-draft funnel Repaired & re-voiced

• Outreach drafts were silently quarantining on a partial output shape — fixed, so a valid draft now reaches review. The orchestrator review path that 500-errored on a re-drafted touch is now an idempotent supersede.
• 9 outreach drafts were regenerated in the new voice and are flowing to the approval queue.
• New copywriter rule: no targeting-rationale openers ("given your role/segment…"), and a brief, recipient-first Misraj intro so a cold contact knows who is reaching out — context that earns the pitch, not a sender-centric headline.

Edit-before-approve + a real learning loop Live

• The operator can now edit an AI draft before approving it directly on the approval dashboard, instead of only approve/reject.
• Every edit is captured as a structured learning signal (original AI text vs. the operator's final text) in a new draft_edit_feedback table — the single highest-leverage input for teaching the copywriter what "good" looks like. This closes a gap the 2026-07-04 audit flagged: the learning loop had no outcome linkage.

LinkedIn content auto-publish Live · safety-envelope

• Sara's and Rawan's approved, scheduled LinkedIn posts now publish to their real personal feeds automatically. LinkedIn has no posting API, so this drives the logged-in browser session through the existing MCP (a small additive create_post tool installed on both seats) — no second browser, no re-login.
• Fail-closed by design: off unless explicitly enabled, and every publish must clear a stack of gates — 1 post/day/account, business hours only (Sun–Thu, 09:00–17:00 Riyadh), operator-approved items only, kill-switch/global-pause respected, idempotent (never double-posts). Drafts are never auto-posted.

Personal-account voice + weekly performance review Live

• The content planner was re-grounded on each person's real voice (read live from their accounts). Sara's and Rawan's posts broadened from Baseer-only to the whole Misraj ecosystem (Arabic-AI thought leadership, not a product ad); Norah's X account is treated as an independent Saudi AI enthusiast — broad AI-interest content, not affiliated with the company (operator correction 2026-07-08).
• A Thursday weekly content-performance review now captures per-post engagement. Twitter/X metrics are read for real; LinkedIn per-post metrics are stored honestly as unavailable via the LinkedIn MCP (it exposes no per-post read) rather than fabricated.

Quality & data model

• ~1,400 automated tests passing (up from ~1,202 yesterday), covering the new projection, enrollment, approval-with-edits, and auto-publish paths.
• Database schema is now at migration 046 — adding draft-edit learning capture (045) and per-post content metrics (046) on top of the approval↔AI-draft linkage (042).

What's next

• Watch the first live auto-published LinkedIn posts land cleanly, then loosen the 1/day cap only after several clean posts with no account warning.
• Feed the captured draft-edits back into the copywriter so the learning loop starts improving drafts, not just recording them.
• Work the email-first exec cohort's first approved sends, and progress the manual Deputy-Minister track.

LinkedIn outreach Live

• Connection requests now send from Sara's and Rawan's own LinkedIn seats, capped at 20 invites/day and 100/week per seat, with ban-safety pacing (3–8 minute randomized gaps between invites, max 8 per batch).
• The kill-switch blocking all LinkedIn sends was deliberately turned off today, with an audited database record of the change and who made it.
• First live invites are scheduled from 09:00 Riyadh time today, 2026-07-06.

Inbound CRM pipeline Live

• New CRM leads in Odoo are now automatically researched by an AI agent, turned into a branded proposal deck, and posted back to the CRM as an internal note for sales — no manual work.
• First lead through the full pipeline: Saudi Aramco, with a live, ready-to-share proposal deck link generated automatically.
• Nothing is emailed to the lead directly — the only external action is an internal CRM note visible to the sales team.

Approve-to-send funnel Repaired

• Found and fixed today: approvals were not actually arming anything to send — a broken link between "you approved this" and "the system sends this" meant zero approved messages had ever gone out. That link is now rebuilt and verified end-to-end.
• Every seat now sends from its own identity — sara@trybaseerocr.com and rawan@trybaseerocr.com — with a consistent bilingual sign-off, replacing a shared generic address.

Safety audit Passed

• A full audit of every live flow — LinkedIn, email, inbound, PR, and the agent runtime itself — ran today, producing a 16-fix hardening wave shipped the same day (pacing bugs, stuck-task recovery, duplicate-send protection, and more).
• 1,202 automated tests passing (up from roughly 1,072 a week ago), plus new process-supervision watchdogs that keep the agent runner, scheduler, and LinkedIn bridge alive and self-restarting.

The outbound sales engine (Phase 1)

• Sourcing — discovers GCC large enterprises in document-heavy industries (banking, insurance, government, telecom, healthcare).
• Qualification — a deep ICP rubric scores every account/contact and disqualifies SMEs, off-segment, and non-GCC targets automatically.
• Journey — per-contact state machine: LinkedIn invite → acceptance → your-approved message → 3-business-day reminders → email hand-off.
• Booking — Cal.com bookings close the loop and stop all outreach to that contact.

The demand & brand engine (Phase 2)

• Signals Hub — Twitter/X, LinkedIn, GitHub, community and news signals in one actionable stream.
• Market intelligence — 7 competitors under watch (ABBYY, Nanonets, Upstage…), research papers, tenders, news.
• Content calendar — weekly planning + publish-on-approval for Sara & Rawan (LinkedIn) and Norah (X).
• PR workspace — 20 seeded GCC press contacts, pitch drafting, approval, reply tracking.

The two runtimes

• Hermes agent runtime — runs the LLM reasoning for every agent on the shared Hermes/OAuth configuration (no per-agent API keys). A 1-minute relay carries each agent's request to the model and back.
• Cloudflare edge — the API worker (business logic, approvals, execution, a 5-minute maintenance cron) and the dashboard, both close to GCC users.

The data source of truth — Neon

• 48 database migrations define accounts, contacts, journeys, approvals, execution queue, signals, competitors, content plans, PR, draft-edit learning capture, content-post metrics, twitter-reply drafts, signal recency, and audit tables.
• Every agent output, review, approval, send, and state change is recorded — the dashboard is a live window onto Neon, never a guess.
• Typed status constraints and idempotency keys make the pipeline crash-safe.

Agent ↔ Agent (via the Orchestrator)

Agents never call each other. A sourcing agent's approved accounts become child tasks the Orchestrator spawns for contact-discovery; approved contacts spawn enrichment; enriched contacts spawn outreach drafts. Depth is capped (≤3), duplicates are deduped, and only whitelisted task types can be spawned — so no runaway loops.

Agent ↔ Neon

The runtime claims a queued task from Neon, gives the agent its data + tools, and records the typed output back to Neon — where it is schema-validated. A valid output moves the task to awaiting orchestrator review; an invalid one is quarantined. Heartbeats and a stale-task reaper keep the 24/7 loop crash-safe.

Agent ↔ Dashboard (you)

When an agent proposes something that sends — a message, reminder, tweet, or pitch — it becomes an approval item you see on the dashboard. You approve, reject, or request an edit. Only your approval, matched exactly to the drafted content, lets the execution layer send it. The dashboard also shows agent health, the task board, quarantine, relay status, and the full contact journey for every prospect.

The four prospecting sources

• Apollo — primary structured account + people search.
• Web search — open-web discovery of companies and Vision-2030/digitization signals.
• LinkedIn MCP — GCC contacts, gated on genuine profile activity (dormant profiles disqualified).
• ScrapeGraphAI — deep, Arabic-aware extraction of evidence from a candidate's own web presence.

Every finding, from any source, must clear the ICP rubric before it enters the pipeline.