ONSET

CORTEX · OPERATIONAL INTELLIGENCE

Ten stages between question and answer.

CORTEX is the operational intelligence pipeline that every ONSET AI Employee runs on. From a customer message to a verified, brand-safe action — ten checkpoints, no hallucinations slipping through.

Most “AI chatbots” pipe a prompt straight to a foundation model and pray. CORTEX doesn't. We chain ten specialist stages so what reaches your customer has been through the same controls a senior account manager would apply — and every action is replayable end-to-end.

Talk to Reginald + ONSET teamView Pricing

MODULE 2 · CORTEX

Ten stages between question and answer.

CORTEX is the operational intelligence pipeline that every AI Employee runs on. From a customer message to a verified, brand-safe action — ten checkpoints, no hallucinations slipping through.

Most “AI chatbots” pipe a prompt straight to a foundation model and pray. CORTEX doesn't. We chain ten specialist stages — Intake, Entity resolution, Collection, Normalisation, Memory, Intelligence, Strategy, Verification, Delivery, QA — so what reaches your customer has been through the same controls a senior account manager would apply.

  1. 01

    Intake

    Capture the message, channel, language, intent

    ingest · ms=42

  2. 02

    Entity

    Resolve who's writing — customer, prospect, supplier

    match · Lim Wei Hao · prospect

  3. 03

    Collection

    Pull prior threads, invoices, contracts, products

    loaded · 14 artefacts

  4. 04

    Normalisation

    Standardise language, dates, currency, names

    BM → EN · RM → RM

  5. 05

    Memory

    Recall preferences, past objections, deal stage

    recall · 6 facts

  6. 06

    Intelligence

    Reason about the actual ask + business context

    reasoning · 1.4s

  7. 07

    Strategy

    Choose tone, offer, escalation path

    strategy · upsell-soft

  8. 08

    Verification

    Fact-check against your data, not the model's guess

    verified · 4/4 claims

  9. 09

    Delivery

    Route to WhatsApp/email/voice; gate at WATCHDOG

    queued · awaiting approval

  10. 10

    QA

    Sample-audit. PULSE writes hash. Loop closes.

    qa · 8.4/10

← swipe through all ten stages →

See a real CORTEX run →

EACH STAGE IN DEPTH

What every checkpoint actually does — and why it's isolated.

Each stage is its own n8n workflow with its own input contract, its own output schema, its own audit row. Failures are bounded to one stage. You can swap any stage's LLM provider, model or even prompt without touching the others.

  1. 01

    Intake

    Capture the message, channel, language, intent

    Without proper intake, every downstream stage is a guess. CORTEX parses channel (WhatsApp / email / voice / web), detects language (BM / EN / 中文 / Manglish / Tamil), and classifies primary intent in ~40ms before anything else runs.

    ingest · ms=42 · channel=whatsapp · lang=ms-MY · intent=quote_request

  2. 02

    Entity

    Resolve who is writing — customer, prospect, supplier

    A "hi can come or not" from a 5-year customer ≠ the same message from a cold lead. Entity resolution dedupes across phone, email, WhatsApp ID and normalised company name so the downstream stages see the right person — with the right history.

    match · Lim Wei Hao · prospect · last seen 14d ago

  3. 03

    Collection

    Pull prior threads, invoices, contracts, products

    Loads everything CORTEX would need to answer correctly. Prior conversations, current invoice status, products they've bought, contract clauses. Limited to the artefacts strictly relevant to the resolved entity + intent — not a "load everything" dump.

    loaded · 14 artefacts · 2 invoices · 3 prior threads

  4. 04

    Normalisation

    Standardise language, dates, currency, names

    BM "esok" → "tomorrow". "RM4.250" → "RM 4,250.00". "Mr. Lim" / "Mr Lim" / "Lim Wei Hao" → canonical. Without normalisation, downstream LLM reasoning fragments on trivial format differences and Strategy picks the wrong tone.

    BM → EN · RM normalised · 4 entities canonicalised

  5. 05

    Memory

    Recall preferences, past objections, deal stage

    Stored facts you've approved over time. "Prefers WhatsApp not email." "Pushed back on Stripe last quarter." "Decision-maker is the finance director, not the GM." Memory is curated, not auto-grown — every fact has a provenance + a confidence label.

    recall · 6 facts · 0 stale

  6. 06

    Intelligence

    Reason about the actual ask + business context

    This is the reasoning step. A higher-reasoning model by default for client-facing logic, a faster model for high-volume internal work. The reasoning sees: the original message, the resolved entity, the collected context, the memory — and produces a structured answer with explicit confidence + evidence pointers.

    reasoning · 1.4s · confidence=high · evidence=4 obs

  7. 07

    Strategy

    Choose tone, offer, escalation path

    Given the Intelligence output, what should we actually do? Friendly nudge or formal escalation? Discount allowed or not (TREASURY enforces)? Auto-respond or route to a human? Strategy is the layer that maps reasoning to a concrete action plan.

    strategy · upsell-soft · within margin floor

  8. 08

    Verification

    Fact-check against your data, not the model guess

    The LLM proposed an action. Before we ship, Verification cross-checks every claim against your verified data. "Invoice was 30 days overdue" — confirmed via Module W ledger? "Bundle costs RM 2,899" — matches the pricing table? Hallucinated claims fail closed and never reach Delivery.

    verified · 4/4 claims · 0 hallucinations blocked

  9. 09

    Delivery

    Route to WhatsApp/email/voice; gate at WATCHDOG

    The action is queued for delivery on the right channel. If it's a reversible action (an info reply, a calendar booking), it ships immediately. If it's irreversible (a payment, a public post, a refund, a collections escalation), Command's WATCHDOG opens the 30-second Telegram cancel window.

    queued · whatsapp · awaiting approval

  10. 10

    QA

    Sample-audit. PULSE writes hash. Loop closes.

    Every action gets a quality score. 5% of all actions get a full re-trace audit. Patterns that fail QA inform the next-cycle calibration. PULSE writes the hash + the full trace; if anything was wrong, you can replay the exact reasoning chain that produced it.

    qa · 8.4/10 · hash written · trace replayable

A REAL CORTEX RUN

We pointed CORTEX at our own company's website.

marketinglancers.com.my — the founder's consultancy. Eight minutes, end-to-end. The report below is real, unsanitised.

Observations

6

from 5 collectors (crawl/seo/aeo/perf/reputation)

Insights

37

across SEO / branding / market / competitor / lead

Recommendations

8

prioritised critical / quick-win / medium / long-term

Verification flags

16

insights auto-flagged for human review by stage 8

SAMPLE RECOMMENDATIONS (REAL CORTEX OUTPUT)

  • criticalAudit and resolve site crawlability and PageSpeed accessibility issues immediately
  • quick winLaunch a Google Business Profile review acquisition campaign targeting existing clients this week
  • quick winRewrite the homepage H1 and primary CTA to be specific, outcome-driven, and audience-targeted
  • mediumImplement FAQ, Article, and breadcrumb schema across the homepage and key service pages
  • long termCreate and publish an llms.txt file and restructure homepage content architecture for topical authority
elapsed · 10.0 min
total cost · 0.63¢ MY
audit events · 12
Run CORTEX on my domainView Pricing

WHY THIS BEATS A “RAG CHATBOT”

RAG is stage 3. CORTEX is stages 1–10.

Vendors selling “AI agents” usually ship Intake + RAG + a single LLM call + Delivery. That's 4 of CORTEX's 10. The missing 6 are why their agents hallucinate.

ConcernGeneric RAG chatbotCORTEX
Entity resolutionTreats every message as a new conversationDedupes across phone / email / WhatsApp / company name (stage 02)
Format normalisationSends raw user text to the LLMBM/EN/中文 + dates + currency normalised before reasoning (stage 04)
Memory of prior commitmentsOnly what is in the promptCurated, provenance-tagged facts with confidence labels (stage 05)
Fact-checkingTrusts the LLM outputCross-validates every claim against verified data (stage 08)
Approval gate on irreversible actionsShips whatever the LLM saysRoutes through Command's WATCHDOG 30s cancel window (stage 09)
AuditabilityLogs the request + response, maybeHash-chained full-trace replay per action (stage 10 + PULSE)

BEFORE YOU BUY

The six things buyers ask about CORTEX.

  • Why ten stages instead of a single LLM call?

    Single-LLM-call systems are pure prompt → response. They hallucinate because they have no fact-checking step (Verification), no business context loader (Collection), no memory of past commitments (Memory), no governed delivery (Delivery + WATCHDOG). CORTEX isolates each concern so we can audit, debug and replace any single stage without rebuilding the rest.

  • How is CORTEX different from RAG?

    RAG is one technique inside Collection (stage 3). CORTEX is the full operational pipeline — Intake, Entity resolution, Normalisation, Memory, Strategy, Verification, Delivery + QA are all separate from "retrieve documents and feed to LLM." Most RAG implementations skip Verification entirely; that's why they hallucinate.

  • What does a real CORTEX run cost?

    A typical run is < RM 0.10 in LLM cost — a fast model at intake/normalisation, a higher-reasoning model at intelligence/strategy/verification. The platform-enforced 60% margin floor means your bundle pricing is set so CORTEX always runs profitably. A per-tenant cost dashboard is available on the Enterprise tier.

  • Can I run CORTEX on my own infrastructure?

    Yes, on the Enterprise tier. The pipeline is portable — Supabase + n8n + OpenRouter is the public-cloud default; we can deploy to your VPC with your own Postgres + your own LLM provider routing. Talk to our team for scoping.

  • How do you handle LLM provider downtime?

    Multi-vendor routing with automatic failover. Our error handler auto-retries up to 3 times across providers with short waits between. If all three fail, the affected job is quarantined (not lost) and re-queued after a cooldown. Zero data loss; we have not lost a customer-impacting job to provider downtime in production.

  • Can I see a real CORTEX report?

    Yes. We ran CORTEX against marketinglancers.com.my (our own consultancy). The report — 37 insights, 8 prioritised recommendations, AEO score, full audit trail — is available as a sanitised PDF on request. WhatsApp our team and we will share it directly.

Stop hoping the AI got it right.

Ten verified checkpoints. One replayable audit chain. Zero hallucinations reaching your customer.

WhatsApp our team