arosplatforms (@arosplatforms) header
arosplatforms (@arosplatforms) profile picture

We build AI OS for businesses.

Toronto, Ontario
Joined June 2026
Photo by arosplatforms (@arosplatforms)Photo by arosplatforms (@arosplatforms)Photo by arosplatforms (@arosplatforms)Photo by arosplatforms (@arosplatforms)Photo by arosplatforms (@arosplatforms)Photo by arosplatforms (@arosplatforms)Photo by arosplatforms (@arosplatforms)Photo by arosplatforms (@arosplatforms)Photo by arosplatforms (@arosplatforms)Photo by arosplatforms (@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
1
5
442
In automotive supply, a defect that escapes the line costs far more at the customer's plant than at the camera. We build computer-vision inspection systems for parts makers designed around the inspector, not around replacing them. The architecture: cameras and lighting engineered per station, models trained on a defect taxonomy agreed with quality engineering, and a confidence band. Clear passes and clear fails move on. Anything uncertain or near a critical feature goes to a human with the image, the region of concern and similar past cases. Every inspector decision is captured and fed back into the next training cycle, so the system learns the plant's own standard. Models are versioned, re-validated on a held-out set before any change, and the station can fall back to manual inspection at any time. What it changes: inspectors spend their attention on the hard cases, and quality has a searchable record of every call. What to measure: escape rate to the customer, false rejects, share of parts routed to review, and time to recognise a new defect type. #Automotive #ComputerVision #AI
3
Hosting data in Canada is necessary. It is not the same as sovereign AI. When we design for Canadian public sector, federally regulated institutions and firms that must keep control of their stack, residency is only the first requirement. We also ask who holds the keys, which legal regime applies to the operator, whether weights and logs can leave the perimeter, and what happens if a vendor changes terms or withdraws a model. The architecture follows those answers. Inference and retrieval stay in an approved region. Keys remain with the client or a Canadian custodian they appoint. Model files, prompts, traces and eval sets are treated as assets under the same inventory and exit plan as any other critical system. A fallback path is documented before go-live, so a change in a frontier API does not halt a regulated process. This is not legal advice. It is a delivery standard: residency, key control, auditability, exit and fallback, written down and tested. Sovereignty is an operating model, not a checkbox on a cloud map. #CanadaTech #SovereignAI #AIGovernance
4
The expensive mistake in mainframe modernization is writing the new system before anyone can explain the old one. We start by reading the estate. Agents work through COBOL, JCL, CICS and copybooks, job schedules and the interface contracts around them. They draft a business-rule catalogue cited to the code, map data lineage, and flag dead paths and hidden batch couplings. That catalogue is the brief for the rewrite, not a substitute for it. Engineers and business owners confirm each rule. The agent changes nothing in production. Only then do we generate candidate services, test harnesses and cutover maps, under human review. Slices go live with parity tests against the incumbent jobs. The mainframe remains the system of record until those tests hold. What it changes: modernization stops depending on a few people who remember how the batch works. What to measure: share of critical jobs with a cited rule catalogue, defects found before cutover, parity failures in parallel run, and time from confirmed rule to live slice. #Mainframe #LegacyModernization #EnterpriseAI
2
Microsoft announced seven new MAI models at Ignite this week, spanning coding, cyber, image, speech and transcription, plus Decision 1.1 for routing. A wider catalogue is useful. It is not a reason to change a production agent. Here's how we treat it in our builds. Vendor benchmarks stay a hypothesis until they are replayed on the client's golden set with the same tools and the same definition of a correct outcome. Price-performance claims get the same treatment: we measure cost per correct decision, including review and rework, not cost per token. A coding model gets tested on the client's repositories, not a leaderboard. A routing model gets compared with the routing rules we already own. Microsoft's own post is clear that some of these models are tuned for everyday tasks, not the hardest long-horizon work. That distinction matters. We pin the model that wins the client's work, and the rest of the stack (permissions, evals, audit, fallback) stays unchanged. More models. Same evaluation bar. #EnterpriseAI #AIModels #LLMOps
27
In diligence, the risk is not reading too slowly. It is a material claim buried in a footnote that never reaches the investment committee. We build M&A due diligence agents that work the data room the way a careful associate should: every finding cited, nothing asserted without a source. Documents are classified by type and permission. Financials, litigation, contracts, employment terms and change-of-control clauses are extracted with page-level citations. A workplan agent then drafts issue lists against the deal team's checklist, flags inconsistencies across documents, and lists what is still missing. It does not value the company. It does not recommend a bid. Partners and counsel decide. The agent is permission-aware, so clean-team folders stay inside their walls, and every answer opens back to its page. What to measure: share of checklist items with a cited answer, missed-item rate on a labelled sample, time to a first issue list, and how often a finding is withdrawn because the citation was wrong. #LegalTech #PrivateEquity #GenAI
1
6
We attack every agent before it is allowed near a live decision. Red-teaming is not a late penetration test. It is a design step. Before go-live we write an abuse and failure catalogue with the business owner: prompt injection through documents and tools, permission escalation, data exfiltration and the errors the process cannot tolerate. Then we run it: - Automated attacks against the agent's tools, retrieval and prompts, including hostile content hidden in ordinary files. - Human sessions by people who were not on the build team. - Access tests that confirm the agent can only see and do what the signed-in user may see and do. Findings become concrete controls: tighter tool allowlists, output filters, stronger abstention, or a decision that the use case is not ready. We do not treat a clean demo as evidence of safety. We treat a surviving attack catalogue, with residual risk accepted in writing, as the gate. What to measure: attacks attempted and blocked, residual high-severity findings, and time to patch a new class of failure. #AISecurity #AIagents #EnterpriseAI
18
A concentrator can lose recovery in a single shift and spend the next week arguing about why. We build mill-optimization agents that advise the plant, not run it. The stack sits read-only on historian and lab data. It watches grind size, throughput, reagent dose, slurry density and recovery together, flags when the plant is drifting from a similar historical regime, and drafts a short recommendation with the evidence behind it. Constraints come from the operation: equipment limits, metallurgical targets, environmental permits and what the shift is allowed to change. The agent never writes a setpoint. Supervisors and metallurgists accept, amend or reject, and every override is kept so we can see where judgment beat the model. What it changes: the conversation on the shift moves from anecdote to a cited comparison with a near neighbour in the data. What to measure: recovery and throughput against the baseline regime, time from drift to a decided action, recommendation acceptance, and whether any advice would have breached a hard limit (it must not). #Mining #IndustrialAI #AI
4
Most retail markdown decisions are still made in spreadsheets, a week too late. We build pricing and markdown agents for retailers that keep the merchant in charge and make every recommendation explainable. Here's the architecture: - A demand layer that learns price elasticity per SKU, store cluster and season, with uncertainty shown, not hidden. - An inventory layer that knows weeks of cover, inbound receipts and end-of-season targets. - A constraint layer for brand rules, price ladders and margin floors set by the business. - An agent that drafts the weekly markdown plan and writes a plain-language reason for each move. Merchants approve, edit or reject. Nothing reaches the price file without sign-off, and every override is captured so the model learns where judgment beat it. What changes: decisions happen weekly instead of at season end, and they are consistent across categories. What to measure: sell-through against plan, gross margin return on inventory, end-of-season residual stock, and the share of recommendations merchants accept. #RetailTech #PricingStrategy #AI
5
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol this week. Here's how we're bringing them into our enterprise builds. We don't swap models because a launch post says so. Every client agent we run has a pinned model, a golden set of real tasks and a cost-per-correct-outcome baseline. A new model earns its place by beating that baseline on the client's own work. So this week we're replaying our golden sets against both: - Astra on the hard, lower-volume steps: multi-document reasoning, exception analysis, changes to legacy integration code. - Sol on the high-volume steps: extraction, classification, first drafts. What we measure: accuracy on labelled cases, abstention when evidence is missing, malformed tool calls, output tokens per task, and full cost per correct decision including human review. OpenAI (@OpenAI) says both models finish demanding work with fewer output tokens. If that holds on client data, routing economics change. If it doesn't, the pinned model stays. Launches are hypotheses. Evals are the decision. #EnterpriseAI #AIModels
20
The best enterprise AI use cases are the boring ones. The demos that win attention are creative: generated campaigns, talking avatars, open-ended assistants. The use cases that pay back are usually quieter. Matching documents. Explaining exceptions. Assembling evidence. Drafting the first version of something a specialist will check. Boring work has properties that make AI succeed. It is high volume, so small gains compound. It has clear rules and known right answers, so it can be evaluated. It has an accountable owner, so someone can approve the outcome. And the cost of today's process is already understood, so the value is easy to see. Open-ended work has none of those advantages. It is hard to measure, hard to govern and hard to defend at budget time. Our advice to leadership teams: start where the work is repetitive, rule-heavy and painful. Build the evaluation, governance and integration foundations there. Then take on the ambitious use cases with something that already works. #AI #AIStrategy #EnterpriseAI
9
Freight invoices are one of the most quietly expensive documents in a supply chain. Carriers bill against contracted rates, fuel surcharge tables and accessorials: detention, liftgate, re-delivery, residential delivery. Rates and contracts change, and shipments rarely match the plan. Most companies audit a sample and pay the rest. This is one of the agent systems we build for shippers and logistics teams. The agent reads every carrier invoice, matches it to the shipment record, proof of delivery and the contract version in force on the ship date, and recalculates what should have been charged. Each variance comes with a plain explanation and the evidence: the rate line, the surcharge table, the delivery timestamps. Accounts payable and transportation teams decide what to pay, short-pay or dispute. The agent never pays or disputes on its own. The same data shows which lanes and carriers drive exceptions, ahead of the next negotiation. What to measure: invoices audited, variance identified versus recovered, and dispute cycle time. #Logistics #SupplyChain #EnterpriseAI
1
20
We do not build enterprise AI on copies of production data sitting in a developer sandbox. It is one of the first questions a security team asks, and it should be. Agents need realistic data to be built and tested. Customer records, claims, patient notes and contracts are exactly what should not be spread across environments. Here is how we handle it. Development runs on masked or synthetic data that preserves the structure and edge cases of the real thing: the odd formats, the missing fields, the contradictory documents. We generate hard cases on purpose, because real data rarely contains enough of them. The golden set used for final evaluation is built from real cases, but it stays inside the client's controlled environment, with access logged and limited to named reviewers. Logs and traces are scrubbed before they leave production, and retention is set with the privacy team, not by default. Good test data is not a compliance tax. It is how we find failures before customers do. #AISecurity #DataPrivacy #LLMOps
11
Trade finance still runs on paper logic, even when the paper is a PDF. Under a letter of credit, specialists check invoices, bills of lading, insurance certificates and certificates of origin against the credit terms and against each other. A spelling difference, a date outside the window or an inconsistent description can be a discrepancy. The work is time-bound and hard to staff. This is one of the agent systems we build for banks. The agent extracts every relevant field from the presented documents, compares them to the credit and to one another, and drafts a discrepancy list with the exact clause, document and field behind each finding. It also screens parties, vessels and ports against the bank's sanctions data and flags matches for review. Document examiners make every decision on compliance and payment. The agent never releases funds or sends notices. What to measure: examination time per presentation, discrepancy agreement with examiners, and missed discrepancies in seeded tests. #FinTech #TradeFinance #AIagents
3
In Canada, an enterprise AI system that only works well in English is not finished. Federal institutions have official languages obligations. Quebec's Charter of the French Language sets expectations for French in commerce and the workplace. And many national organisations simply serve customers and employees in both languages every day. Here is how we build for that. We evaluate French separately. A model that scores well in English can produce French that is grammatically fine but wrong in register, terminology or legal meaning. Our golden sets include French cases written and reviewed by fluent subject matter experts, not machine translations of the English set. We check retrieval across languages, so a French question finds the English policy and the answer cites it correctly. We keep approved bilingual terminology in a managed glossary that the business owns. And we report quality by language, so a gap is visible before users find it. This is a design requirement, not legal advice, and we confirm obligations with each client's counsel. #CanadaTech #AIGovernance #GenAI
6
When yield drops in a semiconductor fab, every hour spent finding the cause is expensive. The evidence exists. Wafer maps, inline metrology, tool sensor traces, maintenance logs, recipe changes and lot genealogy. The problem is that it lives in separate systems, and engineers stitch it together by hand under pressure. This is one of the AI systems we build for advanced manufacturing. When an excursion is detected, the agent classifies the wafer map signature, traces the affected lots back through their process history, and finds what they have in common: a chamber, a tool, a recipe revision, a consumable change or a maintenance event. It produces a ranked root-cause brief with the supporting data for each hypothesis. Yield engineers decide what to hold, what to test and what to change. The agent has no write access to tools or recipes. Confirmed root causes are captured, so the next similar excursion starts from the answer. What to measure: time to root cause, lots exposed per excursion, and repeat excursions. #Semiconductors #Manufacturing #IndustrialAI
4
Tencent released Hy4 flash yesterday, and its pricing changes how we think about long documents. Tencent describes a mixture-of-experts model with 224 billion total parameters, about 12 billion active, multimodal input and a context window of up to one million tokens, priced at roughly US$0.07 per million input tokens. At that price, putting an entire contract set or a full claim file into one prompt starts to look cheap. The question we test is whether it is reliable. Long context is not the same as long recall. Before we let any model replace retrieval in a client build, we run needle tests on the client's own documents: specific clauses, figures and exceptions buried deep in large files, scored on whether the answer is correct and cited to the right page. We also check where the model is hosted and under what terms, since that decides whether it can touch regulated Canadian data at all. Cheap context is an option worth testing. Retrieval stays until the evidence says otherwise. #LLM #GenAI #AIagents
1
22
Water utilities lose a large share of treated water before it ever reaches a meter. Finding where is the hard part. Leaks hide underground. The evidence is scattered across district meter flows, pressure loggers, acoustic sensors, work order history, pipe age and material records, and customer complaints, all held in different systems. This is one of the AI systems we build for municipal and regional water utilities. We model expected night flows per district, detect deviations that suggest new losses, and combine them with pressure and acoustic signals to narrow down the likely location. The agent then produces a ranked investigation list for field crews, each item with the evidence behind it and the asset history of the pipes nearby. Operations managers decide where crews go. The system makes no changes to pumps, valves or pressure settings. Over time, confirmed finds feed back into the model and into capital planning for pipe renewal. What to measure: time from leak onset to detection, confirmed finds per crew day, and district loss trends. #Utilities #SmartCities #AI
9
Every agent we ship comes with a plan for the day it is switched off. Models get deprecated. Vendors have outages. A data feed breaks. A regulator asks a question and the business owner pauses the system while it is answered. If the process cannot run without the agent, the agent has become a single point of failure. So degraded mode is part of our definition of done, not an afterthought. For every agent we document a pause switch that a named business owner can use without engineering, the manual procedure the team falls back to, and the queue that holds work in flight so nothing is lost or processed twice. We also keep the human skills alive. Reviewers handle a share of cases end to end each month, so the team can still do the work when it matters. Then we rehearse it. Before go-live we pause the agent in a controlled window and confirm the fallback actually works. Resilience is not only an infrastructure property. It is an operating model decision. #AIagents #OperationalResilience #EnterpriseAI
9
A security operations centre does not have an alert problem. It has an investigation capacity problem. Every alert needs the same first twenty minutes: who is the user, what is the asset, what else happened around it, has this pattern been seen before. Analysts repeat that work hundreds of times a shift, and the real threats wait in the same queue as the noise. This is how we build SOC triage agents. The agent picks up each alert, enriches it from identity, endpoint, network and threat intelligence sources, correlates related events into a single case, and writes an investigation brief with a draft timeline and a proposed severity, citing every log line it used. It runs read-only. Containment actions such as isolating a host or disabling an account stay with the analyst, behind an explicit approval. We red-team the agent itself too, because attacker-controlled data flows straight into its context. What to measure: time to triage, escalation accuracy against analyst decisions, and missed true positives in seeded tests. #CyberSecurity #AIagents #SecOps
1
14
Insurers leave money on the table after the claim is paid. Most of it sits in subrogation. A faulty appliance. A contractor's fire. Another driver at fault. The recovery opportunity is often written somewhere in the adjuster notes, photos or reports, and it is easy to miss under claim volume. This is one of the agent systems we build for property and casualty carriers. The agent reads every closed and open claim file, flags recovery potential with the specific evidence that supports it, identifies the likely responsible party, and checks limitation dates so nothing ages out quietly. For each referral it assembles a recovery pack: cause of loss, cited documents, paid amounts and the gaps a recovery specialist still needs to close. Recovery specialists decide what to pursue. The agent never contacts third parties or sends demands. What we design for: fewer missed recoveries and earlier referrals. What to measure: referral rate, referral-to-recovery conversion, and recoveries lost to limitation dates. #InsurTech #AIagents #Insurance
8