Deploying Cortex to capture operational knowledge and build compounding intelligence for heavy equipment service operations in Chile.
Every field-based business makes decisions under the same constraints: incomplete information and limited decision-making capacity. The operator doesn’t know which technician is closest. The dispatcher doesn’t know what parts are on the truck. The manager doesn’t know the real cost of running each asset.
The information exists — in someone’s head, in a WhatsApp thread from weeks ago, in a photo that was never labeled. Experience helps, but it’s trapped in one person. It doesn’t transfer. It doesn’t scale. It doesn’t survive turnover.
These problems are industry-agnostic. Heavy equipment, mining, construction, agriculture, logistics — the pattern is the same: people, assets, tasks, and decisions made under uncertainty. The solutions are universal. Only the specific properties change.
This pilot applies that universal intelligence to one specific case — a heavy equipment rental operation in Chile. The problems we solve here transfer to any business where work happens in the field and knowledge lives in people’s heads.
The operation as it runs today. WhatsApp groups, verbal handoffs, spreadsheets maintained by one person that are outdated the moment they’re created. Knowledge trapped in heads. Decisions based on partial information.
Same team, same WhatsApp groups, same workflows. But now every interaction is captured, enriched, cross-referenced, and made permanently retrievable. Intelligence compounds with every message.
By rolling out in weekly phases, each stage acts as its own A/B test. We measure the delta introduced by each new capability in isolation — capture vs. no capture, enrichment vs. raw data, proactive questions vs. passive collection, summarization vs. raw relay. All code is built and tested on local simulations before each phase deploys live.
Messages, photos, voice notes, GPS coordinates, timestamps. Zero friction — the team uses WhatsApp exactly as they already do.
Serverless event-driven pipeline. Every interaction is captured, deduplicated, and structured in real time — messages, media, metadata — before anything reaches the brain.
We’re partnering with Claude as the intelligence core of Cortex. Not just calling an API — we’re engineering context. Using Claude Code for rapid development, tool use to give Claude access to our database and external sources, and agentic loops for multi-step reasoning: induction (generate possibilities), deduction (eliminate branches), abduction (pick the best explanation), and reflection (loop or commit).
Claude applies business context to raw data — extracting equipment identity from plaque photos, enriching messages with inferred metadata, identifying data gaps, scoring message importance, and generating operational summaries. The goal: a system that understands the current state of the business and makes informed recommendations as context compounds.
Properties over columns — every fact stored as a typed row with provenance, confidence score, and evidence trail. The schema never changes when tracking something new. Nothing is ever hard‑deleted.
Context-aware summaries delivered to managers. Priority-scored alerts, daily operational state, and decision-support recommendations — delivered right back through WhatsApp.
Cortex doesn’t just learn for one company. First-principle rules learned from one operation benefit every operation. A principle about bearing failure at altitude, learned in Chilean mining, helps a construction company in Bolivia without anyone making a phone call.
Every new tenant makes the system smarter for every existing tenant. Every message processed, every pattern recognized, every principle abstracted — the knowledge compounds across the entire network. This is the defensibility: the system gets better the more businesses use it.
Rental Patagonia is the first node in this network. Everything we learn here — about equipment, repairs, scheduling, customer patterns — becomes foundational knowledge that benefits every company that joins after.
Each week introduces one new capability on top of the last. Each phase is its own A/B test — we measure the delta of each capability in isolation before stacking the next.
Deploy Cortex relay on active WhatsApp groups. Every message, photo, and voice note is captured and structured in real time. Team onboarding with zero behavior change — they keep using WhatsApp exactly as before.
Activate Claude-powered enrichment. Photos are analyzed for equipment metadata — make, model, year extracted from plaque numbers. Messages tagged with GPS coordinates, timestamps, and contextual labels. Live operational state monitoring.
Cortex begins identifying data gaps and proactively asking clarifying questions through natural conversation. Missing fields, ambiguous references, and incomplete records are resolved in real time.
Replace live relay with intelligent summarizations. Each message is scored for importance and business context. Managers receive structured operational briefs instead of raw message streams.
Full results, validated metrics, and performance analysis published here. Pilot conclusions inform production deployment strategy.
SME assumptions and industry benchmarks. These are our best estimates before live data collection begins. Each metric will be validated and updated weekly as real-world evidence accumulates.
| Metric | Before Cortex | After Cortex |
|---|---|---|
| Data capture rate | <5% Informal, nothing systematically recorded | ~95% Automatic from existing WhatsApp |
| Equipment ID accuracy | ~30% Outdated CSVs maintained by one person, never verified, stale on creation | ~85% AI extraction from plaque photos, cross-verified |
| Knowledge retention | ~20% Lost with employee turnover | 100% Permanently stored, indexed, cross-referenced |
| Manager situational awareness | ~40% Manual check-ins throughout day | ~90% Real-time state + intelligent briefs |
| Query response time | Hours/days Relies on tribal knowledge and phone calls | <30 seconds Instant retrieval from knowledge store |
All estimates are subject-matter-expert priors. Updated weekly with observed data as the pilot progresses.
Every message processed, every enrichment applied, every question asked, and every summary generated is logged with full metadata. Pipeline health, latency, and accuracy metrics are computed continuously.
End-of-week review with the operations team. Qualitative feedback on system utility and operational impact. Quantitative comparison against baseline metrics.
Each weekly phase introduces exactly one new capability. This isolates the effect of each feature — we can attribute changes in metrics directly to the capability that caused them.
All code is tested on local simulations of real conversation flows before live deployment. Live results validate simulation accuracy and inform the next phase’s deployment parameters.
Full pilot analysis, validated metrics, and architecture deep-dive will be published here upon completion.