Conneqt Platform
Brands across the Gulf are flying blind on the half of their business that lives outside their tenant. Conneqt is the layer outside the tenant — a platform that listens to everything beyond the firewall, reasons about it through a structured brand brain, and activates the insight through alerts, audits, content, and conversation.
The pitch is simple: connect a brand, and Conneqt tells you which competitors moved last night, which of your sites are invisible to generative AI, where your customer data is fragmenting across channels, and what you should do about it by tomorrow morning.
The math: 159 sites × 90 days = 14,310 audits per quarter, automated, vs 636 hours of analyst time the manual workflow would consume ($47,700 replaced cost). Plus everything the manual workflow can't do — daily monitoring, regression alerts, GEO scoring, cross-portfolio rollups.
Three structural problems in the Gulf intelligence stack:
Copilot and ChatGPT Enterprise live inside the customer's M365 tenant. They have no view of the open market. Competitor moves, regulatory tracking, social signals, generative-AI citation behaviour, public-market intelligence — all of it sits outside the tenant. Most operators have never measured how much of their decision-relevant data they cannot see.
Customer Data Platforms across the region have matured — Cenomi Plus, Carrefour's MyCLUB, Nahdi's Nuhdeek, AlFuttaim Share — and the data they harvest is rich. But the intelligence layer above the data is missing. Operators have lakes of first-party loyalty signals and zero capacity to turn them into strategy without three vendors and a quarter of integration work.
The MENA market is flooded with flashy AI tools. Almost none of them run daily. Almost none produce audit-grade output. Almost none survive an enterprise security review. The gap between a demo that wows and a platform that operates is where most engagements collapse.
Conneqt collapses these three problems into one loop: ingest → reason → activate. Ingest from outside the tenant and from the first-party CDP layer. Reason through a structured brand brain. Activate through every channel an operator already uses.
A Multi-Tenant Streamlit Application
The platform is a Python-native Streamlit application served on port 8501, structured as a multi-page app via st.navigation, with a custom dark-sidebar design system and a portfolio-first information architecture.
Pages built (eight subscription-gated modules + portfolio overview + subscriptions admin):
Portfolio Overview — Homepage. KPI strip (active clients, modules, wiki articles, sites under audit, portfolio P0s), industry-mix donut, module-coverage matrix, time-saved widget showing automated-vs-manual analyst-hour math, per-client cards with last-fetch / last-audit freshness pills.
Brand Brain — Per-client conversational intelligence. Force-directed Obsidian-style knowledge graph (pyvis) over the brand's wiki layer plus all raw signals. Claude Haiku reasoning over the most relevant wiki articles. Tag cloud across the brain. Recently-updated timeline.
Website Audit — Daily site-level Tech + SEO + GEO health monitoring. SEMrush-equivalent KPI strip (Site Health, critical issues, broken links, speed grade) plus the Beyond-SEMrush row (GEO score, pages crawled, P0 today). Score-distribution histogram, by-brand and by-market rollups, per-site drill-down with sub-page audits and broken-link lists. Trend-over-time chart that accumulates value as the daily fetch runs.
GEO Readiness — Generative-engine optimisation scoring. Six-dimension radar (llms.txt presence, AI-crawler permissions, schema richness, content depth, FAQ structure, authorship signals). Per-site dimension heatmap. AI-blocked sites surfaced for immediate action. Sites missing llms.txt listed for prioritisation. To our knowledge, no commercial tool offers this scoring in the GCC.
Position Tracking — Keyword ranking across markets. Daily SERP capture with day-over-day movers, featured-snippet ownership, by-market rollups.
Backlinks Intel — Referring domains, DA over time, toxic-link detection across the monitored sites.
Brand Benchmarks — Cross-brand × market heatmaps for overall health, GEO, and P0 alerts. Brand-level rollup table. Sub-brand registry rendering each Champion brand's sub-brain. The portfolio platform view for franchisors operating dozens of brands.
Content Engine — Multi-agent content production in the brand's voice (phase 2).
Alerts Feed — Live and historical alerts. P0 sites, unreachable sites, regressions vs prior run, improvements. Drill-down per alert with full audit JSON.
Subscriptions Admin — Operator-side panel to flip modules on/off per client.
The CDP Layer — Where Customer Data Becomes Brand Strategy
Conneqt's Customer Data Platform component sits beneath the dashboard, unifying signals from sources operators already own and from the open-market signals operators don't:
First-party loyalty ingestion — Mall-wide loyalty platforms (Cenomi Plus-style), pharmacy loyalty (Nuhdeek-style), retail receipts and POS feeds, e-commerce platforms (Salla, ZID, Shopify, WooCommerce, Magento). Identity resolution across these sources.
First-party CRM ingestion — Salesforce, HubSpot, Dynamics where present. Mapped to the same unified customer profile.
Behavioural events — Page views, add-to-cart, purchase, search, app sessions. Stored as an event stream and rolled up nightly into customer feature vectors.
Open-market signals — Competitor pricing, brand mentions, sentiment, regulatory changes, generative-AI citation behaviour, SERP positions, backlinks. Fed through the same data spine.
Activation outputs — Feeds the Brand Brain (for reasoning), the Alerts Feed (for triage), the Content Engine (for personalised activation), Telegram broadcasts (for the merchant or operator), WhatsApp Business broadcasts (for the end-customer base).
The architectural decision is to treat the open-market data and the first-party loyalty data as one corpus. Most CDPs handle the first half. Most market intelligence tools handle the second. A Cenomi marketer planning a Decathlon padel campaign for Vision-2030 mass-participation needs both — who's buying gear today, who's about to churn, what competitors just promoted, and how ChatGPT will answer when a Saudi runner asks for advice tomorrow morning.
The Brand Brain — An Obsidian-Style Wiki, Rendered Live
Every client has a brain directory with three subdirectories:
wiki/ — Curated knowledge written and maintained by AI. One markdown file per topic. Tagged (#competitor, #regulation, #vision-2030, etc.), linked via Obsidian-style [[wikilinks]], dated with Last updated footers.
raw/ — Unprocessed source material. Daily fetch logs, competitor data dumps, regulatory feeds, social-listening rollups, paid-media exports.
outputs/ — Generated analyses. Every multi-part answer the brain produces gets saved as a dated file. Brain reasoning is auditable end-to-end.
For portfolio franchisors, the brain supports a sub-brain pattern — each brand under management gets its own scoped CLAUDE.md, INDEX, wiki, raw, and outputs. The master brain reasons across the portfolio; the sub-brains reason inside each brand. Cross-references between layers are explicit.
The Site Auditor
A custom Python auditor runs daily across every site in the client's sites.json. It scores three dimensions:
Tech — HTTPS, security headers, response time, page weight, mobile viewport, Core Web Vitals via PageSpeed Insights API.
SEO — Title, meta description, H1, JSON-LD schema, sitemap, robots.txt, hreflang, canonical, OpenGraph, image alt coverage.
GEO — llms.txt presence, AI-crawler permissions in robots.txt, valuable schema patterns (Organization, Product, FAQPage, BreadcrumbList, AutoDealer, Vehicle), content citability signals (named entities, dates, numbers), semantic structure.
Each audit produces a structured dict with a 0–100 health score per dimension and a list of issues tagged P0 / P1 / P2. Daily runs are concurrent (ThreadPoolExecutor, max 5 workers), with delta detection against the prior day. Regressions of more than 10 score points trigger Telegram alerts.
The Daily Fetch
Per-client scheduled tasks fire each morning. The fetch is a Claude-driven workflow:
Web search across configured topic areas (competitor names, regulatory triggers, sector keywords, brand mentions).
Save raw findings to raw/news/news-YYYY-MM-DD.md with sources cited inline.
Map signals to affected wiki files. Patch them with dated update blocks.
Run the site auditor. Compare to yesterday. Emit deltas.
Push critical findings to the Telegram operations bot.
Print a summary.
The fetch is idempotent, resumable, and runs in under six minutes for a typical portfolio.
The Telegram Operations Bot
Each client has a Telegram bot running on its own token. The bot answers natural-language questions over the brand brain in Arabic and English. It also receives the alerts feed — P0 issues, regressions, unreachable sites, daily-fetch summaries — so an operator can run their morning standup from a phone.
The AI Layer
Powers the dashboard's Brand Brain Q&A and the Telegram bot. Fast, cheap, good enough for most queries.
Runs for richer multi-source synthesis, strategy documents, and the daily-fetch wiki patches.
For brands with data-residency requirements (KSA PDPL, UAE PDPL), sensitive analyses run on a local model with no API egress.
Multi-Tenancy
Every client is a tenant. Data isolation is enforced at the filesystem level — one folder per client under ~/brains/<client_id>/. Each tenant's wiki, raw signals, audit results, and outputs sit inside its own directory tree. The Streamlit shell loads the active client from a JSON registry (clients.json), reads only that client's paths, and renders only the modules that client subscribes to.
Subscription gating is module-level: a client can pay for Brain + Audit + GEO + Alerts (the "Full Brain" tier) or for the smaller Audit + GEO Starter, or for the Brand Benchmarks portfolio tier. The sidebar adapts to the subscription. The admin panel is a single Streamlit page; tier changes propagate immediately.
The platform's users are operators and analysts, not designers. Streamlit lets the same Python engineer who built the auditor build the visualisation. There is no API boundary between the data layer and the UI; the dashboard reads JSON directly. Iteration speed is measured in minutes, not days.
Markdown is portable, version-controllable, human-editable, and plays well with Claude context. Operators can fork a wiki article, edit it in their text editor, and the brain reflects the change on the next reload. A database would have offered nothing in return except operational overhead and lock-in.
The registry of clients fits in a single file. Audit results are append-only daily snapshots. The git history is the audit trail. Zero ops overhead, zero migrations, zero schema drift.
Filesystem permissions are well-understood by every operator. There are no schema collisions to manage, no row-level-security policies to maintain, no tenant-mixing bugs to chase. For an agency-operated platform with dozens of tenants and a handful of operators, this is the right boundary.
GEO Readiness scoring does not exist commercially in the Arabic-speaking market. Daily-cadence audit-at-portfolio-scale via commercial tools is prohibitively expensive (six-figure annual contracts for the volume Conneqt runs nightly for free). Custom gives full control over the scoring formula, the issue taxonomy, and the data we surface to the brain.
Operators check Telegram before they check anything else. The dashboard is for analysis; the Telegram bot is for triage. Meeting operators where they already are is not a feature, it is the product.
Saudi PDPL and UAE PDPL are real. A brand can choose, per analysis type, whether the work runs on Claude (richer output, network egress) or on local Llama (data residency guaranteed, output good enough for most descriptive synthesis).
Knowledge that does not refresh decays into noise. The wiki is treated as a living artefact — every morning, the daily fetch patches the articles that the night's signals affected. The brain at 7 AM is meaningfully different from the brain at 7 AM yesterday.
Most CDPs handle loyalty and POS. Most market-intel tools handle competitors and sentiment. Conneqt's bet is that the brand operators of the Vision-2030 era will need both in the same workflow, with the same identity resolution, in the same reasoning layer. That bet is the centrepiece of the platform.
| Layer | Technology |
|---|---|
| Frontend | Streamlit · Plotly · pyvis · Custom CSS design system |
| Backend | Python 3 · requests · BeautifulSoup 4 · lxml · ThreadPoolExecutor |
| AI — conversational | Claude Haiku 4.5 (Anthropic SDK) |
| AI — analytical | Claude Sonnet 4.6 |
| AI — privacy / local | Llama 3.1 8B via Ollama |
| Knowledge graph | pyvis (force-directed) over Obsidian-style markdown |
| Brain layer | Markdown wiki · raw signal store · dated outputs |
| CDP — first-party | Loyalty + POS + e-commerce (Salla, ZID, Shopify, WooCommerce, Magento) · CRM (Salesforce, HubSpot, Dynamics) · behavioural events |
| CDP — open-market | Competitor monitoring · sentiment · regulatory tracking · generative-AI citation · SERP · backlinks |
| Site auditor | Custom Python — Tech + SEO + GEO scoring · P0/P1/P2 issue taxonomy |
| External APIs | PageSpeed Insights · Web search APIs · Anthropic API |
| Scheduling | Cron-style scheduled tasks · per-client morning fetch + audit |
| Storage | Markdown (wiki) · JSON (registry + audit snapshots) · SQLite (optional) |
| Multi-tenancy | Filesystem-isolated per client · JSON registry · subscription-gated module access |
| Activation channels | Telegram Bot API · WhatsApp Business API · email · webhook |
| Operations | Conneqt-operated · daily monitored · Anthropic Build Partner |
Conneqt is not a dashboard. It is not an AI chat. It is a production operations platform that does three things at once: listens to everything outside the customer's tenant, reasons about it through a structured brand brain, and activates the insight through alerts, audits, content, and conversation — all while a Customer Data Platform layer beneath unifies first-party loyalty signals with open-market intelligence into a single, identity-resolved corpus.
The architecture was chosen for correctness in the Gulf market, not for simplicity. Filesystem multi-tenancy over a SaaS schema is heavier to operate but the right boundary for an agency-run platform. A custom daily auditor over a commercial backlinks/SEO tool is more work to maintain but the only path to GEO scoring at scale. A local Llama alongside the Claude API is more moving parts than running everything in the cloud but the only honest answer to PDPL data residency. These decisions were made because the alternative is a demo that does not survive the first quarter in production.
The platform is operated daily for clients across Saudi Arabia, the UAE, and Qatar. It is built and maintained by Conneqt — an AI agency and Anthropic Build Partner — for brands that want to know what their tenant cannot tell them.