DEV Community: Mikhail Savchenko

NVIDIA releases four open Cosmos 3 simulation models

Mikhail Savchenko — Fri, 06 Nov 2026 00:29:03 +0000

NVIDIA published four new open-weight checkpoints for its Cosmos 3 model family, according to a post on Hugging Face dated November 5, 2026. All four are post-trained versions of Cosmos 3 Nano and carry 16B trainable parameters.

What was released

Cosmos3-Nano-Transfer-Auto produces synthetic driving data offline. It can generate synchronized video across up to 11 camera views, fisheye cameras included, and can add LiDAR. Text prompts steer weather, lighting and time of day.
Cosmos3-Nano-Sim-Auto is the streaming version of that model. It produces short chunks of camera and LiDAR output in a few denoising steps, so a simulation can run long and respond to changing controls.
Cosmos3-Nano-Sim-Transfer takes spatial controls from a physics simulator, such as depth, segmentation and edges, plus a text prompt, and renders color video a policy can consume. Physics and ground-truth state stay in the simulator.
Cosmos3-Nano-Sim-Bimanual predicts the next video frames from robot or camera actions. It supports two-arm setups for the YAM and AgiBot embodiments and can be driven by a policy or by joysticks.

Licensing and requirements

The weights use the OpenMDW1.1 license, which permits commercial and non-commercial use under its terms. Running the Bimanual checkpoint requires Linux, an NVIDIA GPU, Docker and the NVIDIA Container Toolkit. NVIDIA says the published setup was tested on one GB200 and that this is not a minimum-memory figure.

Adapting the models

NVIDIA also publishes post-training recipes that cover four stages, from adapting to a domain through distillation to training on the model's own generated history. Teams can adapt Transfer-Auto to their own camera and LiDAR layout, then distill it into a streaming model. The post advises evaluating on held-out episodes, because a short successful run proves the pipeline executes without showing the model has converged.

The post is candid that the hard problems remain open: the next step is to measure where these models support policy learning and evaluation and where they fall short.

What it means

This is a research and engineering release for robotics and autonomous-driving developers. It does not touch the sales, support or operations workflows most B2B teams automate.

Mistral adds managed agents and apps to its Studio platform

Mikhail Savchenko — Thu, 05 Nov 2026 20:29:31 +0000

Mistral AI has expanded Mistral Studio, its platform for enterprise AI applications, alongside the Mistral Large 4 model, which it nicknames Le Chonk. The update adds managed agents and apps, and extends workflows, enterprise context, agent observability and evaluations.

What is new

Managed agents. Developers can build agents for multi-step tasks. An agent can be given skills that encode a team's procedures, sandboxes for running code and subagents for delegating parts of a job. Hooks let developers insert custom logic at set points in a run. Studio operates the runtime, and developers set limits on the tools, data and actions an agent can reach.

Workflows. A workflow defines which steps follow fixed rules, where an agent reasons and where a person decides. Interrupted runs can resume from where they stopped. A team can describe a process in Mistral Vibe and generate, test and deploy a workflow as code that can be inspected and edited. The release adds event triggers from enterprise systems and a visual graph of each workflow's logic.

Apps. Studio offers hosting, authentication and access controls for internal applications. Teams can assemble them from pre-approved, reusable components, which shrinks the amount of new code needing security review. Apps can call agents and workflows and retrieve documents through search indexes.

Developer path. A new Mistral MCP server lets developers work in Mistral Vibe or a coding agent they already use, test locally and deploy to Studio with one command.

Where it runs and how it is observed

Models and the platform can be deployed in a customer's own cloud, VPC or on-premises environment, in Mistral-hosted form, or in a mix. Traces in OpenTelemetry log every stage of an execution, covering model invocations, tool usage, and sign-offs. Evaluations let teams test changes to prompts, models or tools against their own datasets before release. Mistral says the traces and evaluations stay with the customer.

Mistral's own examples

Mistral describes three internal builds. An incident-investigation agent follows the team's runbook, asks for approval before sensitive actions and posts findings to Slack, which Mistral says saves hundreds of hours a month. A product-operations workflow classifies ideas and customer feedback and routes them to the right team, saving about 35 hours a week. A staffing app recommends people for projects and explains the fit.

Named users and partners include ABANCA, Accenture, Alten, BNP Paribas, Capgemini, CMA CGM and Tata Consultancy Services. The hours-saved figures are Mistral's own and have not been independently verified.

Not stated in the source

The page gives no pricing, plan tiers or availability dates for the new features, so those remain to be confirmed with Mistral.

Rent the Tools, Keep the Record: Why a Business Needs Its Own Portal

Mikhail Savchenko — Thu, 05 Nov 2026 17:23:29 +0000

For most of the last twenty years a small business ran on rented software and on one person who held it together. The CRM had the contacts, the inbox had the conversations, a spreadsheet had the prices and the accounting tool had the invoices. When a client asked where their order stood, someone opened four tabs, matched the names by eye and answered. That person was the integration. It worked because a person remembers what the conversation was about and carries it between screens that never talk to each other.

An AI agent cannot do that, for plain mechanical reasons. It knows nothing about yesterday's call until the call is written down somewhere it can read. It learns that an invoice went out only if the payment service tells it. Give it ten services with ten partial APIs and it sees ten partial pictures of the same client.

What an agent needs that a person did without

To finish a piece of work on its own, an agent goes through three steps in a row. First it reads the current state: who the client is, what they bought, what was promised and by when. Then it acts, booking the visit, sending the reminder or issuing the invoice. Then it checks what happened and corrects itself, which only works if the result of its action lands in the same place it read from.

When the state is split, each step breaks. Reading means stitching together exports that spell the client's name three different ways. Acting is limited to what each vendor's API allows, and many let you read but not write, or write to some objects and not to others. Checking fails most often. The confirmation from the payment provider and the reply in WhatsApp arrive in different systems, and nothing links them to the booking that started it all.

A business can see how far it is from this on its own website. In our census of 5,095 live sites on October 26, 2026, 3.7% ran an MCP server, the interface through which an AI agent can do something on a site instead of only reading the page; among sites in Brazil, Argentina, Chile and Mexico the share was 1.6% (INITE census). The rest offer an agent the text of a page and a phone number. The agent readiness statistics break the census down by country and platform.

Every connector brings rules someone else wrote

Integrations look free at first. Zapier moves a lead from the form to the CRM, and a webhook writes the payment into the spreadsheet. A year later the business has a dozen of these links, and each one carries the data model of the service at its far end: how the CRM defines a deal stage, whom the booking tool counts as a client, what the payment provider calls a refund. When one of them renames a field or retires an endpoint, a link stops working without a sound, and the business finds out from an unhappy client.

The cost of these links grows with every service added, and the vendors set the terms. A business cannot change how a rented CRM describes a lease. It can only work around it.

What to keep and what to rent

None of this requires writing your own accounting, telephony or card processing. They are hard, regulated and well served by vendors. What the business has to keep is the record that connects them.

That record covers clients, conversations, bookings, orders, payments recorded as events, the promises made to each client and the decisions taken on them. The rules go with it, such as who gets an answer at night, when a person takes over and what counts as an overdue payment. All of it sits in one system on the business's own domain, where staff and agents read and add to the same history.

Everything else plugs into that record as a device. The payment provider takes the card and reports the result, the calendar shows the free slots, and the accounting tool receives the invoices. Any of them can be replaced by another, because the history was never stored inside it.

Suppose the CRM subscription ended tomorrow. If knowing what you promised each client last month would then depend on an export nobody has ever run, the vendor holds your memory.

Speed comes from the record

The gain shows first in how fast a client hears back. In a study of 1.25 million online leads published by Harvard Business Review in 2011, companies that tried to reach a prospect within an hour were nearly seven times as likely to qualify the lead as those that waited longer (HBR). An agent can answer within minutes at 11 p.m. on a Sunday only when it sees who is writing, what they looked at and which slots are free, with nobody opening four tabs first.

The history you cannot rebuild later

A portal can be built in weeks. Five years of what happened with each client cannot be rebuilt afterwards, because most of it was never written down in one place: why a client left, what discount was agreed on a call, how a complaint was settled in a chat. A business that starts recording this in its own system today will have an agent that knows its clients a year from now. One that waits will start from an empty table and its vendors' exports.

Two of our reviews show what this looks like in a single trade: what eight contractor websites let a homeowner do, and what seven Brazilian real estate agency websites let a client do. INITE builds this record as Portal, a client website where the booking, the payment and the conversation go into one history per client, and where the assistant learns everything it knows from that history. The domain, the content and the client records stay yours, and the agreement says how you take them with you if you leave. On the INITE platform the memory itself is a separate service, Brain, which other products and agents reach over MCP. The agent memory guide goes through what belongs in the record.

AWS details Smart Approvals for Amazon Quick agents

Mikhail Savchenko — Thu, 05 Nov 2026 05:25:05 +0000

AWS has published a description of Smart Approvals in Amazon Quick, a feature that decides whether an agent's proposed action needs a person to confirm it. The source page does not give a release date, so this item covers what the feature does and how it is controlled.

What it does

Smart Approvals compares each proposed action with the user's request. It then approves the action automatically or returns it to the user. It does not add permissions and does not remove human approvals that are already required.

Policy sets the limits

Before any automated review, Quick checks the action against Custom Permissions, Quick Guardrails, connector action permissions and the user's own settings. Denied actions stay denied. An action is eligible for automated review only when the user's setting is "Ask each time" and nothing else demands human approval.

Administrators use Custom Permissions to decide who can use the feature. It is on by default, but profiles with deny-by-default enabled have to allow it explicitly. Users can switch it off for themselves, and that choice covers conversations and scheduled runs. When a user has switched it off, no account-level setting turns it back on for them.

How the review works

The classifier is a separate model. It looks at what the action does, where data goes and who is affected, using the request and a limited view of the conversation. It can also consult the user's knowledge graph, built from connected sources, to see how a recipient relates to the user, for instance whether that person sits in the user's reporting chain. AWS says this content is not used to train models or improve the service.

Examples that should go back to a person include sending content outside the organisation, sending to senior leaders in the reporting chain, changing access, deleting records and starting payments. AWS says the list is not exhaustive. The request-matching step also acts as a check against indirect prompt injection, where text in retrieved content tries to steer the agent.

AWS is open about the limits. Model review can make mistakes, so actions that must always get human approval should be enforced through connector action permissions. Each automatic approval is single-use and covers only the exact action and parameters reviewed. Timeouts and review failures never produce an approval.

AWS gives an example. An account manager asks for a quarterly review brief to go to two named colleagues. A note in the source material says it should also go to an all-sales list of about 2,000 people. The send is designed to come back to her, because the audience has changed.

Audit trail

Permission updates linked to Smart Approvals are recorded in AWS CloudTrail, with the user, connector, action and time. The marker does not prove the grant was accepted or the action ran, so the destination system, such as the sent email or updated ticket, is the proof of execution. CloudTrail Event history keeps events for 90 days, so longer investigations need a configured trail. Chat logs capture only activity after delivery to Amazon CloudWatch Logs is set up, and they can hold sensitive content, so access should be restricted and data protection policies considered.

Suggested rollout

AWS recommends confirming connector action permissions first, setting up logging, and limiting the feature to a pilot team and a defined workflow. Define success beforehand as fewer prompts for routine actions with consequential ones still reaching a person, and expand only when pilot evidence supports it.

Fal raises $1.3B at a $16B valuation, led by Bessemer

Mikhail Savchenko — Thu, 05 Nov 2026 04:56:48 +0000

AI infrastructure company fal has closed a $1.3 billion Series F round, as reported by Tech.eu. Bessemer Venture Partners led the round, with NVIDIA, Sound Ventures, HOF Capital and T. Rowe Price also taking part.

The numbers

Fal announced the Series F together with a $350 million Series E that Sequoia Capital led in April. The two rounds add up to $1.65 billion. Citing Bloomberg, Tech.eu says the new round values the company at $16 billion, twice the $8 billion valuation from April. Fal was founded in 2021 by Burkay Gür and Görkem Yurtseven.

What fal does

The company builds infrastructure for generative media. Businesses use it to run image, video and audio models at scale and to fit generative AI into the production and creative workflows they already have. The source names entertainment, advertising, e-commerce, gaming and marketing as the main customer industries.

During 2026, fal has worked on the quality and speed of generative media models using post-training techniques. It also continues to develop its own models next to those supplied by partners.

Where the money goes

According to the source, the funding will pay for expanding fal's inference infrastructure and team, and for further work on its generative media technology. The company plans to put money into research and test new approaches to AI-assisted content creation through 2027.

What the source does not say

The article gives no new product, model, price or plan, and no customer terms. Any effect on the cost or speed of generating media for fal's users is not stated and remains unconfirmed.

Why it is on our radar

Generative media is moving from experiments to a layer that companies build production work on, and the investor list in this round points to continued capital for the compute underneath it. For a business that buys such capacity through an API, the practical questions concern vendor dependence and portability rather than the valuation itself.

Meta buys Conductor; its cloud service closes November 20

Mikhail Savchenko — Thu, 05 Nov 2026 04:55:14 +0000

Meta has acquired Conductor, an AI coding startup, according to TechCrunch, which says Axios reported the news first. Meta confirmed the deal to TechCrunch but declined to share further details, so price and terms are unknown.

What Conductor does

Conductor lets a software engineer start a project by putting several agents to work in separate, parallel workspaces, each handling its own part of the job. The agent teams can run in the cloud, on a local machine, or through a phone. The agents are not Conductor's own; they come from major labs such as Anthropic and OpenAI. The company was founded in 2024 and shipped its core product only last summer.

What changes for current users

Founder and CEO Charlie Holtz announced that Conductor is joining Meta. He said the core features will keep working as they do today until early next year, and that the team plans "something new" for customers by then. What that product will be has not been described, so treat it as unconfirmed.

The exception is the cloud service. It shuts down on November 20, and Conductor will then start deleting user cloud data. Anyone with work stored there has a short window to move it.

Why Meta wants it

TechCrunch describes Meta as a company that has spent recent months catching up in AI services. In August it launched its own coding agent, Muse Code. Meta also reported that its Muse agents have passed 10 million users. The Conductor deal adds to a run of AI acquisitions over the past year, including the agent platform Moltbook.

Holtz said he is looking forward to working with Alexandr Wang, Meta's chief AI officer, and Nat Friedman, head of product at Meta's Superintelligence Labs.

What to watch

The source does not say whether Conductor will keep supporting agents from Anthropic and OpenAI after the move to Meta, or whether the next product will be tied to Meta's own models. Until Meta says so, plan on the assumption that today's setup is temporary.

Dharma-AI releases 2B Field Extractor for document-to-JSON

Mikhail Savchenko — Wed, 04 Nov 2026 23:06:00 +0000

Dharma-AI published details of Field Extractor on Hugging Face on November 4, 2026. It is a 2-billion-parameter model for pulling specific fields out of documents. Plain OCR returns all the text on a page and leaves someone to write code that maps each string to a database field. This model returns structured JSON instead.

Two ways to ask

The model accepts a document image plus one of two request types:

Open: no schema is supplied. The model works out from layout and context which parts are fields, such as a label followed by a value or a table with a header row, and returns key-value pairs.
Schema-directed: you specify the fields you need, assigning each a type (string, number, or boolean), using the key names already in place in your system. The output follows that shape exactly, with no extra keys. A field that is not on the document comes back as null instead of an invented value.

Tables are filled row by row. An invoice, for example, returns its line items as an ordered list, each with quantity and value, next to the totals. According to Dharma-AI, one checkpoint handles every document type with no per-type fine-tuning.

How it was built

The base is DharmaOCR-LITE v1.1, a new 2B version of the company's OCR model released in April, already tuned with supervised fine-tuning and DPO. On top of it the team trained a LoRA adapter on 43,384 examples covering both request modes. The data has three streams: English documents collected from the public web, real Brazilian documents in Portuguese sourced internally, and a smaller multilingual share in other European and Asian languages. All were annotated at field level and cover receipts, invoices, forms, bank statements and ID documents, among others.

What the benchmark says

The benchmark has 546 documents and 1,638 cases across three scenarios: all fields with no schema, a partial schema of real fields, and a schema that also includes fields that do not exist. The score combines field-level F1, value similarity on matched fields and the share of valid JSON outputs.

Field Extractor scored 0.834. The four frontier configurations ranged from 0.816 (Claude Opus 5) to 0.741 (GPT-5.6 Luna). Dharma-AI puts its cost at roughly 6 to 147 times lower, depending on the comparison model.

The two sides are priced differently. Frontier costs are based on measured tokens per page at list prices, with reasoning set to low. The $2.24-per-hour rate Dharma-AI cites is for renting one L40S GPU for an hour. The company says this is not a sales price and that self-hosted cost varies with hardware, volume and latency. Dharma-AI also reports tests on two public datasets, SROIE and FUNSD, that were held out of training. All of these numbers come from the vendor itself, and no independent replication is cited.

What is available

The post links a demo for trying the model on your own documents. The source does not state licensing terms, weight availability or a commercial offering for Field Extractor, so those remain unconfirmed.

Together AI puts agent-tuned Compiler on production endpoints

Mikhail Savchenko — Wed, 04 Nov 2026 21:06:12 +0000

Together AI has announced Together Compiler, a system in which AI agents continuously optimize how open models are served. The company presents it as the first step toward what it calls Polymorphic Inference: serving that adapts to new model architectures, shifting traffic and changing chips instead of relying on one fixed configuration.

What Together Compiler does

Together describes the tool as a loop that treats inference tuning as a search problem for agents instead of a classic compiler translation task. A swarm of agents works against a persistent engine that supports fast iteration. According to the company, it runs in production today on some of its most popular serverless endpoints and has found further gains on code that was already heavily optimized. The company calls the practice Continuous Compilation. Previously a new workload meant new hand-written kernels; the claim now is that the system adapts to new traffic shapes automatically.

The source gives no speed or cost figures for these gains, so the size of the improvement is unconfirmed.

Why Together says static setups fall short

The post argues that typical open-source serving recipes ship with one tuned configuration, which may not fit another setting. It cites mixture-of-experts kernels that perform well on random tokens but may not carry over to real traffic. It also notes that 80% of OpenRouter traffic goes to models released in the last three months, and that the ten most popular open models each use a different architecture. Together says its own earlier work on megakernels taught it that fusing everything into one kernel is the wrong goal for this environment.

What is available and what comes later

Together Compiler is available to end customers in a limited beta; access is by request.
Soon, customers are meant to optimize dedicated endpoints for their own workloads at the click of a button.
Other items on the research list: speculative decoding that adapts per request, a router for sending requests across different deployments, post-training of open models against a customer's own tools, and workload-aware routing across several models.

No dates, prices or plan names are given for any of these.

Practical reading

The announcement is mainly a statement about how Together intends to keep open-model serving fast as models turn over quickly. Teams that run production workloads on open models there should treat the beta as an option to evaluate, and should test against their own traffic, since the claims are the vendor's own and come without published benchmarks.

Agentforce Coworker gets $50 per-user plan and Teams access

Mikhail Savchenko — Wed, 04 Nov 2026 20:05:36 +0000

Salesforce has announced an expansion of Agentforce Coworker, the assistant that takes a job from an employee, coordinates the relevant Agentforce agents and workflows, and returns results and suggested actions for review. The company says more than 185,000 people now use it and that usage grew 21x in three months.

What is available now

Four things are live today:

Skills. A workflow that has proven itself becomes a short request anyone on the team can run. Salesforce's example is a top rep's renewal playbook becoming the team standard.
Conversation History. A user can resume a conversation across Salesforce, Teams and the new apps with context intact.
Google Workspace and Microsoft 365 connections. Email, calendars and files around an account are combined with Salesforce data, so a meeting brief or renewal plan covers more than CRM records.
A per-user plan at $50 per user per month. Salesforce frames it as a way to start small, measure value and widen use. It notes that pricing and packaging may change.

Customers can turn Coworker on in Setup. Giving every user a seat on the per-user plan goes through a Salesforce account executive.

What comes later

In November 2026 the company expects general availability of Coworker in Microsoft Teams, an iOS app, desktop apps for Mac and Windows, an MCP server for connectivity, and a free trial for Flex Credits customers. In December 2026 it lists Daily Digest, scheduled tasks, and Coworker in the Field Service, Consumer Goods Cloud and Life Sciences Cloud mobile apps. Salesforce adds that availability varies by region and that unreleased features should not drive purchase decisions.

Governance

Salesforce says Coworker works within existing employee permissions, sharing rules and access controls. Administrators can manage access to supported agents and actions, and approval and audit capabilities exist for supported operations.

Customer figures

These come from Salesforce's own announcement and have not been independently checked:

Fulton Bank runs more than 20 use cases for about 3,000 users.
The Adecco Group is rolling Coworker out across 40 countries.
BACA Systems enabled it in 20 minutes in an environment with 1,300 custom objects. It reports that questions which used to take 45 to 60 minutes of switching between systems were answered directly.

What to watch

The usage growth figure has no stated baseline, so it says little about absolute scale. The customer examples are large organisations. A smaller team should judge the product on its own pilot, using one repeatable workflow and a before-and-after time measurement.

Anthropic previews Claude Duet, an always-on background agent

Mikhail Savchenko — Wed, 04 Nov 2026 19:19:32 +0000

Anthropic has released Claude Duet in research preview, according to a Zapier Blog hands-on by Wade Burrell, a Zapier product marketing manager. Zapier sells automation that plugs into Duet, so read its comparisons with that in mind. Duet is a separate app from the Claude chatbot, though you sign in with your existing Claude account.

What it does

Instead of waiting for a prompt, Duet sends a morning brief summarising the most important items across connected apps. It checks those accounts every few hours and messages you only when something needs attention. You choose what it reports and how often. The work runs inside its own virtual machine that you reach through a browser, meaning your phone or laptop can be switched off while it continues. That browser also lets it fill in forms on sites that have no connector.

The app has three tabs: Chat (text or voice), Tasks (background work) and Make, which holds templates such as a daily to-do list and email triage. Tools, projects, memory and skills from your Claude account carry over. Onboarding pushes four first-party connections: Gmail, Slack, Calendar and Google Drive.

Safeguards

Anthropic states that Duet can read, write drafts, and get work ready without asking first, yet it presents its plan and requires sign-off before sending a message, issuing a payment, executing a document, or removing any data. Trading securities and moving money between accounts stay off-limits even with approval. Duet connects through OAuth, so it does not see passwords or sign-in tokens, and any app can be disconnected. Anthropic also reports prompt-injection screening, a classifier for sensitive data and testing by internal red teams and an outside firm. The reviewer's advice is to check what it does, especially early on.

Access and cost

Access is by waitlist for U.S. Pro, Max and Ultra subscribers, on iOS, Android and claude.ai. There is no separate price. Duet shares your plan's limits, and background checks count against them. The reviewer's first couple of days used up all of his session limits, and the app did not make clear which limits applied.

Early findings

After a few days, the reviewer found setup quick and the proactive behaviour useful. Overnight, Duet matched Google security alerts to the connections he had just made and summarised them unprompted. The Tasks tab starts empty and offers to scan your tools, and two tasks were running by that morning. The downsides were the fast limit use, a Gmail connector limited to one account, and the general risk that an agent misjudges urgency. These are a few days of first impressions on a preview product, not a verdict.

Bedrock adds India-only processing for Kimi K3

Mikhail Savchenko — Wed, 04 Nov 2026 06:00:50 +0000

AWS has announced in-country inference for Kimi K3 on Amazon Bedrock, so requests to the model can be processed entirely within India. Kimi K3 is a multi-modal, open-weight model from Moonshot AI, which reported frontier-level coding and general agentic performance at release.

Two profiles, two trade-offs

Kimi K3 is served through cross-Region inference profiles. You call the model from a source Region, and Bedrock routes the request to a destination Region. The profile ID you invoke decides which destinations are eligible:

Global (global.moonshotai.kimi-k3): AWS recommends it for most uses, as the lowest-cost and most widely available option.
India geographic (in.moonshotai.kimi-k3): for workloads where processing has to stay in the Mumbai and Hyderabad Regions.

AWS says traffic between Regions is encrypted and stays on the AWS network, with no need to build multi-Region VPC infrastructure. The model card is the place to check which source Regions can call each profile.

What you need to change

The post uses the Mumbai Region as the source and shows an IAM policy that allows the model in both Indian Regions plus the India inference profile. Organisations that restrict AWS Regions need to keep the source Region permitted and allow routing to both Indian destinations. The post points to a profile-scoped exception where that fits. The model is reachable through OpenAI-compatible Chat Completions and Responses APIs, as well as Converse and InvokeModel.

One detail matters for audits: metrics and logs are stored in the source Region where the request was made, whichever destination handled it. CloudTrail data events can record the destination Region.

Document workloads as the example

AWS demonstrates the profile with a scanned Hindi electricity bill. The model reads the image directly, extracts fields as JSON under a schema, writes an English summary and answers follow-up questions in Hinglish, with no separate OCR or translation step. AWS advises verifying extraction quality for your own use case and adding validation or human review for critical workflows.

Cost levers

The post lists a flex service tier at a 50 percent discount on standard pricing for latency-tolerant work, and a priority tier at a higher price. Prompt caching, implicit or explicit, discounts repeated prompt prefixes. The post does not say whether the India profile is priced differently from the global one, so check the pricing page.

AWS quotes Ultrahuman, whose health assistant handles more than 14,000 conversations a day. The company reports an 80 percent cut in LLM costs and 85 percent faster multi-turn chats with prompt caching. These are customer-reported figures, not independent benchmarks.

CyberKord closes €2.7M pre-seed for AI-run IT security ops

Mikhail Savchenko — Wed, 04 Nov 2026 05:59:20 +0000

Italian startup CyberKord has closed a €2.7 million pre-seed round, Tech.eu reports. 360 Capital led it, with Plus Partners, Addendum Capital and Vento also taking part. The company was founded in Milan in 2026 by Alessandro Maggio, Dragos Tanasa and Cosimo Michelagnoli.

What the platform does

CyberKord targets the gap between spotting a problem and fixing it. Security tools can flag a threat within minutes, the company argues, but resolution drags because people have to step in by hand. The platform is meant to take over that manual work.

According to the report, it starts from a security alert, a support ticket or a message sent through Microsoft Teams or Slack. It works out the underlying cause, simulates the proposed fix against a Digital Twin of the company's IT infrastructure, and then applies the change using tools the company already runs.

Control and integration

CyberKord says it connects to both cloud and on-premise systems, so companies do not have to replace their current stack. Customers decide how much autonomy the AI gets, and each action is recorded and open to review. CEO Alessandro Maggio frames the aim as freeing expert staff from patches, access requests and one-ticket-at-a-time incidents so they can return to architecture and projects.

What comes next

The money will go toward further product development and the first pilot projects. Those pilots are planned with large European companies in manufacturing, technology, healthcare and telecommunications. CyberKord also plans to hire, mainly in technical and commercial roles.

What is not in the announcement

The source gives no pricing, no availability date and no customer results. The platform is described by its makers, and its claims are not independently verified. Nothing indicates a version for smaller companies yet.