Foundation · what the backbone tokenizes
Every signal the field twin reads.
Operational, logistics and asset signals from the systems the operator already runs. Hover or tap any signal to see where it comes from.
Industry · Oil & Gas
An intelligence layer on top of the systems the operator already runs — SCADA, historians, CMMS, ERP, well systems and the spreadsheets in between. Models that anticipate non-productive time, optimizers that plan the day at the lowest cost, and agents that act before the frac set stops. Nothing replaced; everything connected.
01 · SharpAI
Predictive models
10 models · foundational + downstream
02 · AUTONOMY
Agents
8 agent patterns in this vertical
03 · FDE
Engineering services
2 anonymized engagements
The value chain
In unconventional operations, one hour of non-productive time can cost as much as weeks of freight. The decisions that prevent it are spread across completions, logistics, maintenance and procurement — and today most of them are made by hand.
01
Upstream · completions
Keep the frac set pumping. Anticipate the inputs, crews and equipment each stage needs before they become the bottleneck.
Decisions we model
02
Field logistics
Turn the daily dispatch from an Excel exercise into an optimized plan: how many orders, from which quarry, with which truck and carrier.
Decisions we model
03
Assets
Read pump, compressor and rotating-equipment signals continuously and plan the intervention before the failure.
Decisions we model
04
Supply chain
Spare parts, imports and suppliers planned against the operation's real demand — not against last year's averages.
Decisions we model
05
Downstream
Demand-driven assortment and basket intelligence for service-station stores, planned per site instead of per supply.
Decisions we model
01 · SharpAI · Oil & Gas
We start by modeling the operation as it really is — quarries, hubs, PADs, fleets, assets and their constraints — fed by the operator's own data inside its own cloud. Predictive heads and optimizers attach to that shared representation, so every new use case reuses the same foundation instead of starting from zero.
Foundation · what the backbone tokenizes
Operational, logistics and asset signals from the systems the operator already runs. Hover or tap any signal to see where it comes from.
Downstream · what the head consumes
The early-warning head reads the twin's state plus a focused set of decision-time signals, and recalculates the coverage clock continuously. Hover any node — these are the inputs behind one alert.
Foundation models · self-supervised pre-training, frozen backbone reused downstream
A network model of the operation — quarries, transfer hubs, PADs, wellsites, fleets and carriers — with a daily material balance and the real constraints of each node. It is the shared state every optimizer, forecast and agent reads, and the place where what-if scenarios are tested before they reach the field.
Node-network simulation · daily balance + stock alerts · scenario engine · MILP optimization layer
Corpus · Built per operator on its own data — frac and production plans, dispatch records, well systems, ERP and field spreadsheets — deployed inside the operator's cloud (raw → bronze → silver → gold)
A sequence model over sensor time-series from pumps, compressors and rotating equipment. It learns each asset's normal operating signature from the historian and maintenance history, and exposes a condition embedding that failure, maintenance and readiness heads consume.
Sequence encoder over sensor windows · per-asset condition embedding · trained per operator and asset class
Downstream heads · fine-tuned on the frozen foundation embedding per institution and use case
A coverage clock per wellsite: how many hours of operation the current stock and trucks en route can sustain, recalculated continuously, with hours of lead time to act.
Probabilistic forecast over the twin state · continuous recalculation
The lowest-cost daily plan — orders, quarry, truck type and carrier — that still meets the frac schedule, with every constraint made explicit.
MILP over the twin network · cost curve as the benchmark
Forecasts the activity calendar and the demand for proppant, water, crews and equipment it implies.
Time-series forecast conditioned on plan revisions
Arrival-time prediction per truck and load, learned from historical routes, road conditions and carrier behavior.
Regression head over GPS and route history
Remaining-useful-life estimate for pumps and compressors, so maintenance is planned into the schedule instead of interrupting it.
Survival head over the asset condition embedding
Plans critical spares and consumables against forecast activity and failure risk, balancing stock-out cost against inventory cost.
Intermittent-demand forecast + service-level optimization
Predicts arrival and clearance time for imported equipment and materials, flagging the shipments that put the schedule at risk.
Event-sequence model over shipment milestones
Demand-driven assortment and product-combination recommendations for service-station stores, clustered by site profile.
Market basket analysis + collaborative filtering · store clustering
02 · AUTONOMY · Oil & Gas
A score on a dashboard does not prevent a stoppage. These agents watch the twin, decide inside the operator's rules, and act through the channels the field already uses — with a person approving whatever the policy says a person must approve.
8 agent patterns · ordered by deployment maturity
Watches the coverage clock per wellsite and, when it crosses the threshold, proposes the corrective dispatch — which hub, which trucks, which carrier — to the logistics coordinator.
Trigger · Coverage below threshold
↳ Hours of lead time instead of minutes
Runs the optimizer every morning on the latest plans and stock, explains the recommended plan against the reference cost, and publishes it once approved.
Trigger · Daily planning cycle
↳ Lowest-cost plan, explained
Correlates operational, logistics and asset signals to flag emerging NPT risk on each set, and routes the alert to the person who can act on it.
Trigger · Risk score shift
↳ Earlier, targeted escalation
Turns remaining-useful-life estimates into proposed work orders in the CMMS, aligned to the activity calendar and spare-parts availability.
Trigger · RUL below window
↳ Planned instead of unplanned stops
Monitors critical spares against forecast demand and failure risk, and drafts purchase requests in the ERP before the stock-out.
Trigger · Projected stock-out
↳ Critical stock covered
Follows each shipment's milestones, reads and checks the documentation, and alerts procurement when a delay threatens the schedule.
Trigger · Milestone delay
↳ Fewer surprises at the yard
Collects, validates and loads supplier documentation into the procurement system, escalating only the exceptions to the team.
Trigger · New supplier request
↳ Less manual data entry
Answers planners' and supervisors' questions in natural language over the twin — stock, ETAs, scenarios, costs — with every answer tied to its source data.
Trigger · On demand
↳ One live field picture
03 · FDE · Oil & Gas
Senior engineers embedded with completions, logistics, maintenance and supply-chain teams. We start with a quick win, measure its impact against the operator's own reference, and scale from there — inside the operator's cloud, on the operator's data, with know-how transferred to its team.
SCADA, historians (PI), CMMS, ERP (SAP), well systems (OpenWells), GPS and the spreadsheets that hold the rest — connected read-only first, without replacing anything that already works.
A layered data lake (raw → bronze → silver → gold) and an operational database deployed inside the operator's own AWS or Azure account. The data never leaves.
MILP optimization, simulation and forecasting built together, so a prediction becomes a plan and a plan respects every physical and contractual constraint.
Web applications for planners and coordinators — daily plan, what-if scenarios, alerts — designed with the people who will use them every morning.
The AI recommends and explains; the person decides. Approval steps, audit trail and override reasons are part of the product, not an afterthought.
Discovery workshop, assessment, proof of value and implementation. From simulator to digital twin to predictive control tower with agents — one measured step at a time.
Anonymized engagements · same engineering team end-to-end
Upstream operator · LATAM
Unconventional oil & gas operator
Sand logistics orchestration for frac operations: field twin with daily balance and stock alerts, MILP fleet and order optimizer, what-if web app and a data layer inside the operator's AWS. Roadmap to a GPS-fed digital twin, predictive ETAs and an agentic control tower.
Phase 1 in delivery — the daily plan is measured against the operator's own reference cost curve.
Integrated energy company · LATAM
Service-station convenience network
Demand-driven product mix per store with SharpAI: market basket analysis, collaborative filtering and store clustering, delivered with a dashboard and tested in a controlled pilot.
10 of 12 product combinations tested in the pilot lifted sales by more than 45%.
Team shape — typical FDE composition for this vertical
Embedded senior engineers, predictive models in production, and agents that execute on every score.