Full menu of AI components
The twenty-seven components in detail, grouped by family, with what each one solves.
In operation
16
Solutions in production
12
Forecasting models
7
Similarity algorithms
6
Scoring engines
7
Countries with operations
Three platforms of our own sustain this inventory
The components did not come out of a laboratory. They came out of these three platforms, which operate today. They can be licensed whole, but there is no need: what gets deployed at a client are the pieces.
CFO AI
Financial management platform. Extraction from the ERP, multi-entity consolidation with intercompany eliminations, and a forecasting architecture that separates how much from when: the monthly volume comes out of four models competing per account and is distributed by weekly masks learned from that same account.
On top of that, the neurosymbolic view blends the machine prediction with the chief financial officer's guidance —the final figure is the baseline multiplied by its multipliers and added to its additives, each one recorded separately—, variance is aggregated as variance and not as a naive sum of intervals, and the horizon is marked by data zones: hard, mixed and soft depending on the week. All of it runs in procedures in the database itself.
StockSense
Inventory and demand intelligence platform. It classifies each item by its consumption pattern —regular, erratic, intermittent or irregular, according to mean interval between demands and variability—, and that classification decides which model is applied to it out of seven that compete under expanding-window cross-validation.
From there come safety stock by non-parametric resampling, the economic order quantity, redistribution between points with excess and points with shortage, and a causal graph where the statistical drivers and the business constants propagate in topological order down to the derived values. A four-factor confidence score governs what can be executed without intervention and what cannot.
Vamos Eventos
Mass-consumer platform, and the test bench for the ingestion line. It processes heterogeneous public sources in a five-station pipeline —classify, enrich, merge, media, publish— where each station claims its batch without blocking the others.
It is the component where automatic quality control was proven at volume: extraction in a single model call that replaced a chain of four and reduced inference cost by 75%, deduplication by seven combined signals, the collapse of close to two hundred elements into some forty operations before writing, and a nightly audit that reviews a sample of five hundred records with a model different from the one that produced them.
What this section proves and what it does not prove
That these platforms operate proves that the components withstand load, fail in known ways and have someone to maintain them. It does not prove that they will perform the same on another organization's data: that depends on the history available and on the quality of the source system, and it is exactly what a trial against the client's real data measures before committing to a scope.
Documents and knowledge
The pieces that turn paper, files and email into something you can query. It is where a first project usually starts, because the volume is high and the error is cheap to detect.
Document ingestion and classification
Multimodal reading of PDF, image, spreadsheet, CSV and JSON, classification by type with structured output, field extraction and normalization of whatever arrives without a fixed format —every supplier sends it differently and the pipeline brings it to a single schema—. The output is validated against a strict schema: what does not fit is rejected and flagged for human review, never approximated.
What it solves. Eliminates manual keying over a high volume of documents, and the error rate that comes with it.
Cited retrieval
A retrieval-augmented platform on pgvector, built over the client's own documentation. Fragments are indexed as 768-dimension vectors with an HNSW index and retrieved by semantic similarity. Every answer cites the document and the passage it came from. If there is no source, there is no answer: the component prefers to admit that it does not know.
What it solves. Turns a document repository nobody consults into an answer whose backing can be verified by whoever receives it.
Multi-signal deduplication
Seven algorithms over the same pair of records: cosine over embeddings, Jaccard, Levenshtein edit distance, bigram overlap, n-grams, normalized exact match and phonetic search, resolved by a weighted scoring with a threshold configurable per domain. It detects the same object arriving from two different sources, written two different ways.
What it solves. Keeps the consolidation of several sources from duplicating what is in fact a single record, which is the defect that ruins a dashboard before any model error does.
Signal extraction from an external document
A document that arrives from outside —a supplier report, a circular, a third party's table— is read and translated into a proposed adjustment to a model variable, not into a new figure. Untrusted content is processed between explicit delimiters as a defense against instruction injection, and every adjustment proposed by the model enters with a confidence penalty and with its origin attached. Nothing is applied without a person accepting it.
What it solves. Makes it possible to incorporate information that is in none of the client's systems without that information entering the reported number unchecked.
Data, finance and integration
The least showy part, and the one that most conditions the rest. The bottleneck in these projects is almost never the model: it is getting the data out of the source system with enough permission, quality and frequency.
Multi-entity financial consolidation
Operating today in CFO AI. An extraction agent that reads straight from the ERP —SAP Business One, Dynamics, Sage, Oracle EBS, JD Edwards, Epicor, CONTPAQi—, multi-entity consolidation with intercompany eliminations, exception reconciliation, automatic narrative of variances against prior periods, and natural-language queries over the repository. The calculation lives in database procedures, not in the model.
What it solves. Replaces the manual assembly of the close and leaves every figure traceable back to the system it came from.
Treasury and working capital
Payment velocity per customer computed as a materialized view: average days to payment, volatility of that term and on-time compliance rate. On the supplier side, payment lag orders the disbursement priority. On top of that, a thirteen-week cash projection and a sensitivity analysis that moves five assumptions ten percent in both directions and ranks their impact on the closing balance.
What it solves. Turns the question of whether the cash will hold into a range with explicit assumptions, instead of a single number nobody knows how was put together.
Connectors and multi-format ingestion
One per platform, according to what each one offers: API, exports scheduled by SFTP or email, or RPA automation when there is no interface to consume. The adapter pattern incorporates incremental synchronization, rate control, master-data caching and retries; it operates today against an enterprise ERP across fourteen integration points. Integration by API and by MCP with the existing systems. Each new platform is integration work of its own, and its inventory is taken before a scope is closed.
What it solves. Makes it possible to read from systems that were not designed to be read, without modifying them and without touching their operation.
Receivables prioritization
Ingestion of reports in heterogeneous formats, a collections queue ordered by a recoverability score, notification by email and by WhatsApp with delivery logging, and an aging dashboard by thirty-, sixty- and ninety-day brackets, with per-tenant isolation at the database level. The order the component proposes has to be calibrated against the real behavior of the client's receivables before it is used to assign work.
What it solves. Replaces management by pure aging —the oldest first— with an order that also weighs the probability of recovering.
Models on the client's data
The five components that learn from the history or from the client's own signal. The pattern is the same in every case; what changes is the calibration against the organization's data, which is done before putting them into operation.
Computer vision
Person detection and anonymous trajectory tracking over polygonal zones defined per floor plan: heat maps, dwell time per zone and compliance rules with alerts. Anonymous and aggregated analytics, without biometrics, without facial recognition and without emotion analysis, in line with Panama's Law 81 of 2019. It requires per-site calibration —angles, lighting and zones of interest— defined during a visit.
What it solves. Turns cameras that only record into a source of measurement and of alerts.
Demand and volume forecasting
Each series is first classified by its pattern —mean interval between demands against variability of the size— as regular, erratic, intermittent or irregular, and that class determines which models compete. Seven candidates for series of sporadic consumption, including Croston and its variants, exponential smoothing and seasonal harmonic regression; four for financial series. Expanding-window cross-validation over three folds and selection by lowest mean absolute error, with an ensemble of models when it adds over the best individual one, and a confidence boost when the candidates converge.
What it solves. Makes it possible to size against the expected volume instead of against the average, which is where you lose in both directions.
Anomaly and shrinkage detection
Classic machine learning over transactions and operations, with composite risk scoring and dynamic thresholds. Applied to inventory, it measures the deviation between physical stock and logical stock and scales it by age and by location. It is not generative and it does not close cases: it flags and it ranks, and the decision stays with the analyst. It requires a calibration period against the client's real cases to set the threshold.
What it solves. Ranks by risk a volume that today is reviewed by sampling or by a fixed threshold.
Inventory planning and replenishment
Safety stock by non-parametric resampling: lead times are sampled from their distribution and demand is resampled with replacement over ten thousand simulations, taking the ninety-fifth percentile of demand during the lead time; with fewer than three observations the component falls back to a closed-form formula instead of feigning precision. On top of that, reorder point, economic order quantity, per-supplier lead times, and redistribution from points with more than four weeks of coverage toward those in shortage.
What it solves. Replaces the fixed minimum per item, which does not distinguish between a stable lead time and one that doubles from one month to the next.
Predictive maintenance
Urgency scoring per asset over hours or kilometers of operation, deviations from the equipment's own mean interval, failure history and component type, with seasonal modifiers and a projection of remaining useful life. It escalates in priority levels with a configurable threshold. The pattern moved from vehicle fleets to industrial equipment without being rewritten.
What it solves. Brings the intervention forward to the point where it is still maintenance and not replacement.
Decision and orchestration
What separates a demonstration from a system that passes an audit. These five components do not produce content: they fix who decides, by what rule, and how it is proven afterwards that the decision was the right one.
Neurosymbolic architecture
The model translates, the engine decides. A deterministic inference engine over explicit, versioned rules, a console where each rule is drafted, dated and approved, a derivation trace behind every assertion, and a regulatory knowledge base with an effective date. In the graph variant, the variables are connected in a directed acyclic graph and the effect of moving one propagates in topological order down to the derived values, with tracking of the delta between the baseline and the adjusted version. The model does not originate figures: it contributes language, not content.
What it solves. Makes an automatic decision explainable and reproducible years after it was taken, independently of the model that drafted it.
Rule-based scoring engine
Configurable rules with weights, parallel evaluation of all the candidates, a dynamic threshold and manual override by exception with a record of who exercised it. Six different implementations in production over domains that bear no resemblance to one another —assignment of technical staff, collections prioritization, investment scoring, relevance, quality and maintenance urgency—, which in practice means that what changes per client is the table of weights, not the engine.
What it solves. Turns a criterion that today lives in one person's head into a rule that is written down, arguable and auditable.
Agent orchestration
A dispatch engine and agents that execute over the existing systems, governed by a taxonomy of four types —agent to internal human, agent to system, agent to external human, and agent to agent— each with different production requirements according to its risk. In the ensemble variant, several specialized agents with domain tools of their own and a selector that picks which one handles each query. The agent-to-agent level is still early and is not proposed for production.
What it solves. Fixes what an agent may execute without human intervention, before it executes anything.
Systematic model evaluation
Quality control of the model's outputs before production, to choose which one is used, and during production, to detect degradation. It includes routing between providers by task type with a fallback chain: a light model for classification, a heavy model for reasoning. Both of these run today on our own platforms. The set of cases against which degradation is measured is built on the client's domain, and is part of the deployment.
What it solves. Replaces the impression that the system “answers well” with a repeatable measurement.
Adversarial audit of outputs
A model different from the one that produced the result audits it afterwards, over a fixed sample and field by field, and issues a verdict per field instead of a global grade. Where the accuracy of the record falls below the threshold, the record is withdrawn automatically instead of being published with a notice nobody reads. It runs at night, against the day's output, at a marginal cost next to that of the pipeline it audits.
What it solves. Detects silent degradation, which is the real failure mode of these systems: they do not stop answering, they start answering worse.
Operations, channels and infrastructure
Where the answer comes out, who receives it and where the compute runs. These are the decisions a client tends to leave for last and which in practice determine which use cases are even possible.
Conversational channels
A WhatsApp bot in production on the business API, with signature validation on reception, handling of attachments and images, and segmentation of long replies. Live assistance to the human agent while the conversation is happening, and a call summary logged in the CRM on hanging up. The channel handles inquiries; any operation that moves money is confirmed separately and outside the conversation thread.
What it solves. Offloads the repetitive work from the human channel and gives back the after-call time, which tends to be a high fraction of the shift.
Offline field application
A progressive web application with a service worker and local storage: it operates without coverage and synchronizes on reconnecting. Dynamic forms with photo capture, coordinates and signature, validation at the point of capture, and closing of the work order from the device itself. Three applications in this family are in production, field-tested with intermittent connectivity.
What it solves. Eliminates the paper inspection that someone has to transcribe afterwards, and with it the gap of days between the finding and its recording.
Dispatch engine
Fourteen assignment rules —skills, service commitment, proximity and workload— resolved by weighted scoring, with model-assisted reordering for the context the rules do not capture. In operation over more than thirty-four thousand orders a year across seven countries. The rollout is progressive by design: first recommendation mode, with the dispatcher deciding, and only afterwards automatic assignment.
What it solves. Replaces assignment by the experience of whichever dispatcher is on shift with an explicit criterion, the same for every shift.
Notifications and commitment tracking
Templates with variables, delivery by email and by WhatsApp, automatic alerts on status transitions, and escalation by category and severity of the finding. Logging of deliveries, opens and replies. The outbound queue guarantees delivery with retries: a message that did not go out stays pending, it is not lost.
What it solves. Closes the distance between the system detecting something and whoever must act finding out about it.
Self-service portal
The end customer checks status, opens requests and reviews their history without calling. Automatic linking of their equipment or their account by serial number or identifier, so that there is no prior manual registration. It is deployed against the same system of record the internal staff uses.
What it solves. Takes the status inquiry off the phone channel, which is the most frequent reason for contact and the one of least value.
Real-time dashboards
Metrics aggregated in procedures in the database itself and rendered on the server, with automatic revalidation, charts, filterable tables and indicators with trend. Different views per role over the same data. Where it applies, the interface is generated from the model's response as blocks —table, chart, card, panel with the SQL query in view— so that the user can see which query produced what they are reading.
What it solves. Delivers the number and, next to the number, the query that produced it, which is what makes it possible to argue about it.
Infrastructure
Three cross-cutting pieces that do not show in a demonstration and without which none of the above passes a security review.
Sensitivity-based inference routing
Each type of content is routed according to the sensitivity of the data: on-premises processing or a managed model. The open-weight models —Gemma, from Google— run on servers the client designates, so that regulated data does not leave their network and capacity does not remain in the hands of a provider that can change price or conditions of access. It is configuration by type of content, not a later migration.
What it solves. Enables use cases over regulated data that would otherwise fall outside of scope.
Isolation and concurrent work queue
Per-tenant isolation at the database layer through row-level security, with the policy enforced by the engine and not by the application code: a programming error does not open another tenant's data. The work queue claims tasks atomically without locking between workers, supports concurrent execution with locking by resource, and recovers stuck tasks on its own.
What it solves. Lets several units or subsidiaries share one installation without any of them seeing another's data, and lets the volume grow without rewriting the processing.
Observability and proactive alerts
Periodic health checks per service with latency logging, database triggers on the status transitions, and an alert sent through the agreed channel when something goes from healthy to degraded or down, with its corresponding recovery notice. A heartbeat from the worker process to tell a downed service from a downed monitor.
What it solves. Makes the operator find out from the system and not from the user, which is the difference between an incident and a complaint.
The state of these components corresponds to August 2026 and should be reviewed every six months. The operating figures cited correspond to Limestone platforms and to deployments of our own, not to a performance commitment on another organization's data. Regulatory references are indicative and do not constitute legal or compliance advice.
Ready to talk?
If you're evaluating an AI solution for your company or want to explore a partnership, a managing partner will respond within 24 hours.