arosplatforms

I built an AI team with Grok Bot for ArosPlatforms. One request: “Get me ready for today.” But the real question wasn’t what it could do. It was what I should LET it do. Full build + security walkthrough 👇 @bot @SpaceX Grok Bot Explained: I Built an AI Team for My Business youtu.be/SMCgdqjom-0 #GrokBot #AIAgents

"We need clean data first" is the most expensive sentence in enterprise AI. It turns a ten-week build into a two-year programme, and the data is never clean enough. Our view: you do not need clean data. You need a named owner for each source, a definition of what a wrong answer costs, and a way for the system to say "I don't know". So we start with the decision, not the data lake. We pick the narrow set of sources that decision depends on, assign an owner to each, and measure the errors that matter. The agent cites its sources and abstains when they conflict. Corrections flow back to the owner. Data quality then improves where it pays, because the errors are visible and attributed. What to measure: sources with an owner, conflicts surfaced per week, and time to resolve each. Data readiness is a delivery practice, not a prerequisite. #AI #AIStrategy #DataGovernance

Mainframe modernization stalls on one problem: nobody can say what the code actually does. The COBOL runs the business. The people who wrote it have retired. The documentation describes a system from 2004. Here is how we approach it. We start by reading, not rewriting. Agents analyse the code, copybooks, job schedules and data layouts, then draft plain-language descriptions of business rules, each linked to the exact lines. Domain experts confirm or correct them, and each confirmed rule becomes a test case. Only then do we decide what to migrate, wrap or retire, and we compare outputs old against new on real transaction samples. Engineers and business owners sign off on every rule. The agent does not change production code. What to measure: rules confirmed by experts, behavioural differences found before cutover, and coverage of the batch estate. #AI #Mainframe #LegacyModernization

If your document pipeline reads untrusted files, those files can give instructions to your AI. This is the security test we run early in every build, before the agent is connected to anything that matters. We plant hostile content in realistic documents: invoices, emails, contracts, spreadsheets. Some say "ignore your instructions". Some hide text in white fonts or metadata. Some try to get the agent to call a tool, change a payee, or reveal another case's data. Then we check the layers. Does the agent have the permission at all? Do inputs and outputs get filtered? Does a write go through a second control? Can one customer's data surface in another's answer? What to measure: attack success rate by technique, tools reachable from untrusted content, and time to detect and contain. The goal is not a clean demo. It is an agent whose worst day is still contained. #AISecurity #CyberSecurity #AIagents

Pension administration runs on exceptions: a missing service record, a conflicting date of birth, a plan amendment from 1998 that changes the calculation. Here is how we approach case-prep agents for pension and benefits administrators. An administrator opens a retirement or transfer case. The agent gathers the member's history from payroll feeds, scanned records and plan documents, and flags gaps and contradictions. It identifies the plan provisions that apply, with the clause cited, and drafts the case summary and the member letter for review. Administrators decide. The agent never calculates a final entitlement, approves a payment, or contacts a member. What to measure: - Time to a complete case file - Gaps found before calculation instead of after - Letter edits per case - Calculation reversals The outcome we design for is fewer surprises late in the process, for the member and the plan. #AI #Pensions #AIagents

Quebec's Law 25 has a requirement every enterprise AI builder should read closely: when a decision about a person is made exclusively by automated processing, the organisation must inform them. That affects design, not just paperwork. Here is how we handle it in builds for Canadian clients. We map every agent step to a decision type. Where a step could be the sole basis of a decision about an individual, we either keep a human reviewer in the decision or build the notice, explanation and review path in from the start. We also keep logs that show which inputs and which version of the system produced an outcome, so an explanation can be reconstructed later. What to measure: decisions with a documented human or notice path, explanation requests answered on time, and logs complete on audit. This is engineering guidance, not legal advice. Your counsel owns the interpretation. #CanadaTech #Law25 #AIGovernance

Warranty data holds an automaker's earliest warning of a quality problem. Most of it is free text written by technicians. Here is how we would build warranty root-cause agents for a manufacturer. We combine claim narratives, repair codes, parts traceability, build records and field telemetry into one searchable view. The agent clusters claims that look like the same underlying failure, even when technicians describe it differently, and drafts an investigation brief for the quality engineer: suspected component, affected build window, supporting claims, and what is still unknown. Engineers decide on containment, supplier escalation and recalls. The agent never classifies a safety issue on its own. What to measure: - Time from first claim cluster to engineering review - Claims correctly grouped, sampled by experts - False cluster rate - Share of briefs engineers accept as a starting point Earlier detection is the outcome we design for. The evidence decides whether we got there. #Manufacturing #Automotive #EnterpriseAI

A fine-tuned 27B model that holds its base model's score while using 25% fewer tokens is worth testing. It is not yet worth deploying. Continuity Research released Tess-4.5-27B this week, an Apache-2.0 fine-tune of Qwen3.8-27B for agentic coding, built for local hardware. The team reports the same aggregate benchmark pass rate as the base model with fewer tokens. Here is how it would be considered in our builds. Coding agents appear in our work in legacy modernization: reading old code, documenting business rules, and drafting test cases for engineers to review. Token use drives the cost of those loops. So we would test it on a client's own code and tests, in the client's tenancy, and compare cost per accepted change, not tokens per run. What to measure: accepted-change rate, reviewer effort, and tool-call error rate. Vendor numbers choose what we test. Our evals decide what we use. #AI #LLM #OpenWeights

Covenant monitoring in commercial lending is a document problem disguised as a data problem. The inputs are credit agreements, quarterly compliance certificates, financial statements and amendments. The obligations sit in clause text, and the numbers sit in PDFs and spreadsheets. Here is how we would build it for a commercial bank. First, extract each covenant into a structured record: definition, test frequency, thresholds, cure rights, and the clause it came from. Second, recompute each ratio from the borrower's submitted financials using the agreement's own definitions. Third, draft a review note for the credit officer, with every figure traced to a page. Credit officers decide on waivers, watch-listing and escalation. The agent never communicates with borrowers. What to measure: - Breaches caught before the quarterly review - Time to prepare a covenant review - Definition mismatches found - Officer corrections per note #FinTech #CommercialBanking #AIagents

A prompt change is a production change. Most teams still ship it like a text edit. In our delivery method, every agent has a release gate, and nothing reaches production without passing it. The gate has four parts: 1. A regression suite built from the client's own golden cases, including past errors we promised never to repeat. 2. A side-by-side run of the current and candidate versions on the same inputs, scored per field and per segment. 3. A cost and latency check at expected peak volume. 4. A named approver who signs the change record. This applies to prompt edits, model upgrades, retrieval changes and tool changes alike. If a vendor deprecates a model, the same gate decides the replacement. What to measure: regressions caught before release, change lead time, and incidents traced to unreviewed changes. Speed in enterprise AI comes from a gate that is cheap to run, not from skipping it. #LLMOps #AIGovernance #EnterpriseAI

Every vulnerability scanner produces a backlog. The hard question is which findings your business can actually be hurt by this month. Here is how we approach exposure triage for security teams. We join scanner output with what the scanner cannot see: asset ownership, internet exposure, compensating controls, and whether the vulnerable component is reachable in how it is deployed. Threat intelligence adds whether exploitation is being observed. An agent then prepares a ranked remediation worklist. Each item carries the evidence, the proposed owner, and the patch or mitigation path, in plain language the owning team can act on. Security engineers confirm priorities. The agent never patches, changes firewall rules or closes findings by itself. What to measure: - Time from finding to owner assignment - Share of critical exposures remediated inside policy - Rework caused by wrong ownership - Findings suppressed without evidence, which should be zero A shorter list is only an improvement when it is a more correct list. #CyberSecurity #GenAI #AIagents

A telecom network operations centre does not lack alarms. It lacks a fast way to know which twelve of the forty thousand matter. Here is how we would architect alarm correlation for a carrier NOC. We start with topology: services, circuits, sites and maintenance windows loaded as a graph the agent can query. Then alarms and tickets are clustered into probable incidents, with the evidence listed for each cluster. The agent drafts an incident brief: suspected root domain, affected customers or services, the last change near the fault, and the runbook step that fits. It cites every alarm and log line it used. Engineers decide. The agent does not push configuration changes or close tickets. What to measure: - Alarm-to-incident compression - Time to a correct first diagnosis - Share of briefs engineers accept without rework - Missed-escalation rate, reviewed weekly Noise reduction only counts if the real faults still surface first. #AI #Telecom #AIagents

Thinking Machines released Inkling 2 Small this week. Here is where a model like it fits in our enterprise builds, and where it does not yet. It is a 594B mixture-of-experts model with 17B active parameters, hosted on their Tinker platform, and it handles text and images. Per the lab, it is closed-weight at release, with weights planned after further safety testing. What we do with a release like this: 1. Run it on a client-style golden set with masked data, not on the vendor's benchmarks. 2. Pin the exact model version behind our gateway so behaviour cannot change underneath a production agent. 3. Treat "weights later" as a plan to re-test in-tenancy, not as a date to build on. Where it could fit: image-plus-text document steps and tool-calling agents that sit behind human approval. What we measure: cost per correct case, abstention quality, and how the answer changes between versions. A new model earns a role in an architecture only after it passes our evals. #EnterpriseAI #LLM #AIEvals

Demo accuracy is a vanity metric. The cost of being wrong should set the autonomy. That is our point of view, and it decides how much an agent is allowed to do. A 97% field on a vendor slide tells you nothing if the 3% includes a wire release, a coverage denial or a missed allergen. Those errors are not interchangeable, so the thresholds cannot be either. We set per-decision autonomy from the downside: what happens if the agent is confidently incorrect, who can catch it, and how fast. High-cost errors stay recommend-only, with a reviewer and a required citation, until shadow-mode data on that exact segment says otherwise. Low-cost, reversible actions can move faster. One average score cannot govern both. Leaders should ask a simpler question than "how accurate is it?": show me the error the system is allowed to make, and the one it must never make unsupervised. Design autonomy from the blast radius. The model score comes after. #AI #AIGovernance #Leadership

A water-quality excursion is a race against incomplete SCADA trends, lab results and work orders, while operators are already handling the plant. We build investigation-pack agents for treatment and distribution exceptions — read-only on historian, LIMS and maintenance data. When a turbidity, chlorine or pressure signal leaves the band operations has set, the agent assembles the surrounding trends, recent maintenance, related lab samples and similar past events into a cited brief. It proposes likely contributing factors as ranked hypotheses, not commands. It never opens a valve, changes a setpoint or issues a boil-water notice. Operators and process engineers decide. The plant's control system stays untouched by the agent. What we design for: faster time-to-a-shared-picture on an excursion, fewer verbal handovers between shifts, and a packet a regulator or after-action review can replay. If the trend or the lab result is missing, the agent says so and stops. #Utilities #IndustrialAI #CanadaTech

The dangerous moment in an agent system is not the draft. It is the write-back. Behind the scenes, we treat every change an agent might make in an ERP, claims system or document store as a payment: it needs a maker, a checker and a receipt. The agent may propose a structured action — a journal, a ticket update, a flagged invoice — with the evidence pack attached. A named human approves it in a screen that shows the diff, the citations and what the agent was not allowed to do. Only then does a separate, narrowly scoped tool execute the write, with an idempotency key so a retry cannot double-post. We log who approved, what changed, and how to reverse it. Irreversible actions get a second approver or stay out of the agent's reach entirely. Shadow mode earns the right to draft. It does not earn the right to commit. If you cannot show the approval record next to the system change, you do not have control. You have a script with better manners. #AIagents #AIGovernance #EnterpriseAI

Workers' compensation files stall when medical notes, job demands, wage records and prior claims live in different systems, and the case manager rebuilds the timeline by hand. We build claims-intelligence agents for that reconstruction — not to decide entitlement. The agent assembles a cited chronology from the FNOL, clinical reports, job descriptions and correspondence the claims team already holds. It flags missing documents, date conflicts and return-to-work constraints that appear in the file, and drafts a case brief for the adjuster or case manager. It never accepts, denies or communicates with the worker or a provider. Every statement in the brief points at a page. Ambiguity is labelled as ambiguity. Humans own coverage, causation and every outreach. What we design for: faster time-to-a-complete-file, fewer dropped documents, and a chronology a new reviewer can pick up without a verbal handover. If a medical conclusion is not in the record, the agent does not infer one. #InsurTech #AIagents #EnterpriseAI

In Canada, bilingual is not a translation step you add at the end. For Quebec and for many federal workflows, English and French are production languages with equal weight. Here's how we design for that. We build separate golden sets and red-team suites in both languages, including Quebec legal and industry terminology the client owns. Retrieval is tested cross-language: a French query must land on the right English source document, and the reverse, before generation starts. Notices, explanations and reviewer screens ship in the language the user is working in. We do not treat machine-translated evals as a substitute for native review. A model that answers fluently in French but cites the wrong decree is a compliance incident, not a localization bug. Logs, prompts and traces stay in the residency region the workload was approved for. Language does not become an excuse to send data elsewhere. If the system cannot retrieve and explain in both languages to the same standard, it is not ready for that deployment. #CanadaTech #GenAI #Quebec

Maverick spend and duplicate suppliers rarely announce themselves. They hide in free-text invoice lines, expired rate cards and contract clauses nobody has read since onboarding. We build procurement-intelligence agents for category managers — read-only on spend, contracts and supplier files until a buyer acts. The agent classifies spend against the approved taxonomy, extracts rates, term dates and volume commitments from the contract, compares invoices to those terms, and drafts a leakage or renewal brief with cited lines. Supplier-risk flags come from documents the company already holds, not from scraped gossip. The agent never emails a supplier or commits spend. Buyers accept, reject or re-route every recommendation. The tolerances and category rules belong to procurement, not to the model. What we design for: a shorter path from invoice anomaly to a cited conversation, cleaner master data for the next cycle, and a log of which leaks were real. If the contract page is not in the file, the agent does not invent the rate. #Procurement #SupplyChain #AIagents

Nous Research released Arete-27B this week: a supervised fine-tune of Qwen3.8-27B, with an open 803-prompt evaluation that scores factual recall separately from how answers are framed. On their eval, answers they tagged as adopting a state-aligned account fell from 22% on the base model to under 1% after the fine-tune, while their general-capability benches held roughly flat. We are not adopting anyone's political eval as a release gate. The lesson for enterprise work is simpler. A fine-tune can change what a model will assert, refuse or omit without moving SWE-bench. When we evaluate an open-weight checkpoint for a client's gateway, we re-run behavioral tests on that client's suite: framing and refusal, tool misuse, data leakage, and EN/FR parity — not just task accuracy. Vendor and community scores stay in the research folder. Promotion requires the client's golden set, a pinned version and a documented rollback. If you cannot say how a fine-tune changed the model's behavior, you are not ready to route production traffic to it. #LLM #OpenWeights #AIGovernance

Sustainability reports fail in the footnotes: activity data without a source, emission factors without a version, and Scope 3 methods that cannot be replayed. We build ESG reporting agents that assemble an assurance-ready evidence pack — not a marketing draft. The agent extracts activity data from the systems and files the sustainability team designates, binds each figure to a source extract and an emission-factor record with its vintage, labels estimated versus measured values, and drafts disclosure language mapped to the framework the company has chosen. Gaps show up as gaps, not as invented completeness. It never files a report or picks a factor the team has not approved. Every number a reviewer sees can be opened back to the extract it came from. What we design for: fewer last-week scrambles, a lineage an assurer can sample, and a clear list of what the company still does not know. If the factor or the activity record is missing, the agent leaves the cell blank and explains why. #ESG #Sustainability #EnterpriseAI