Extreme automation and elimination of operational bottlenecks.
4 terms · agents, hallucination, integration, data
- A2AAgent-to-Agent
- Direct communication between two or more AI Agents. Instead of a human intermediating the information, a Sales Agent (SDR) converses with a BI Agent to qualify a lead using predictive analytics in real time.
- Hallucination
- When an AI model invents information with great confidence, presenting false data as true. Happens often in generic AIs. The solution is to use RAG and Guardrails.
- APIApplication Programming Interface
- It's the "waiter" of the technology world. The API is the channel that allows two different software to talk to each other without needing to know how the other was built. When your CRM (HubSpot) needs to send lead data to MyDatAgent's AI, it does so through an API.
- Medallion Architecture
- A way to organize data in layers: Bronze (raw data), Silver (cleaned and standardized) and Gold (ready for AI and strategic analysis).
Very high risk for compliance and business decisions.
Integration and automation. It's thanks to APIs that your AI can read and act on company systems (Jira, Slack, ERP) in real time, without manual data copies.
Ensures that your company's AI drinks clean water, without errors.
5 terms · reasoning, command line, infrastructure, scale, security
- Chain of Thought
- The ability of advanced models (like MDA LLM 2.1) to "think step by step" before giving an answer. Instead of guessing the final answer, the AI breaks down the logical problem into stages.
- CLICommand Line Interface
- The traditional and direct way developers and systems communicate with software through plain text commands, without graphical interfaces (buttons or windows). In the MyDatAgent ecosystem, CLI connection is a powerful alternative to using agent protocols (like MCP).
While MCP requires the AI to read complex tool schemas and "decide" what to do (consuming many tokens in the process), CLI allows sending direct and lean data and instructions to the MDA LLM 2.1 — like in automation scripts or batch processing. - Private Cloud
- Cloud infrastructure dedicated exclusively to your company, not shared with the public.
- Cluster
- A group of connected computers (servers) that work together as if they were a single super-powerful machine. Instead of relying on a single server that might crash or be slow, processing is divided among the cluster's machines.
- Content Filter
- The "filter" for security that evaluates everything entering and leaving the AI model. Automatically blocks malicious prompts (Prompt Injection), sensitive data (PII) and responses involving prohibited topics for the company — financial advice, medical advice, profanity. On MyDatAgent they run natively via LiteLLM, without adding latency.
Solving complex business and financial problems with high accuracy.
Drastic token and cost savings. By cutting the context overhead of tools, CLI requests focus only on the task (e.g., summarizing 1,000 contracts, refactoring code). In high-volume operations and repetitive tasks, using CLI instead of agent interactions can reduce token consumption by over 40% — lowering your monthly infrastructure bill without losing model intelligence.
Data sovereignty, zero risk of noisy neighbors and assurance that your data is in Brazilian datacenters (LGPD).
Reliability and scale. Ensures that if your AI operation has a usage spike (e.g., 100 salespeople using the SDR Agent at the same time), the system doesn't crash and doesn't slow down. The load is distributed.
Compliance and brand protection. Ensures that AI doesn't generate illegal, prejudiced or legally risky content — an invisible and non-negotiable bodyguard.
5 terms · data lake, datacenter, dataops, deploy, protection
- Data Lake
- A massive repository that stores raw company data in any format (spreadsheets, texts, system logs, audio), from various sources, without needing to organize everything before storing. MyDatAgent transforms that "lake" into useful information through the BI Agent.
- Datacenter
- The physical "factory" where servers, storage disks and network cabling live that process and store company and internet data. A Brazilian datacenter means that data processed there is under Brazilian jurisdiction, away from foreign laws like the US CLOUD Act.
- DataOpsData Operations
- The discipline and set of practices to ensure that data enters the system, is cleaned, standardized and delivered with quality and security to those who (or what) need to use it — in this case, the AI. It's equivalent to quality control on a factory production line.
- Deploy
- The act of taking software, an AI model or an update that was in testing phase and "putting it live" for daily company use. It's pressing the "publish" button safely.
- DLPData Loss Prevention
- Technologies and policies that prevent sensitive data (like SSN, passwords, source code) from leaving the company network. In the AI context, DLP prevents employees from pasting that data into ChatGPT.
360-degree view. Ends the "islands" of data scattered across different departments. It's the fundamental raw material for AI to generate strategic reports and not hallucinate.
Sovereignty and LGPD. The physical location of the datacenter is what ensures compliance with the General Data Protection Law and protects the company from international corporate espionage.
Reliability of AI. If the data entering the AI is wrong or dirty, strategic decisions will be wrong. DataOps ensures that the AI drinks from a clean and automated source.
Speed of delivery. Automated and safe deployment processes (CI/CD) mean the company can fix errors or launch new AI features in hours, not months.
Protection of intellectual property and LGPD compliance.
1 term · data engineering
- ETLExtract, Transform, Load
- The process of taking data from various places (Extraction), cleaning and standardizing it (Transformation) and placing it in a Data Lake (Loading).
It's the heavy data work that MyDatAgent's DataOps Agent automates, saving months of human labor.
4 terms · fallback, training, quantization, frontier
- Fallback
- The redundancy strategy where, if the primary AI model fails (overuse, internal error, server crash), the system automatically redirects the request to a secondary backup model — without the user noticing the failure.
- Fine-Tuning
- The process of training an already ready AI model with company-specific data (e.g., your contracts, manuals, support history).
- FP88-bit Quantization
- A technique to "compact" a large AI model, reducing memory usage (VRAM) without losing intelligence.
- Frontier Models
- The most advanced and intelligent AI models in the world at any given time (e.g., GPT-5.5, Claude Opus 4.7, MDA LLM 2.1).
Business continuity. Zero downtime. If your SDR or BI operation can't stop, fallback ensures intelligence keeps running on another model until the primary comes back online.
Transforms a generic AI into a business expert, dramatically increasing accuracy.
Reduces GPU infrastructure cost by up to 50%, making MDA LLM 2.1 cost-effective.
Using open-source Frontier Models allows having the world's best intelligence running privately and cheaply.
2 terms · gateway, security
- GatewayEntrance · AI Gateway
- The "gatekeeper" or control center for all network data traffic. In the AI context, the AI Gateway is the point where all AI requests and responses pass. It's where we apply the rules: who can use which model, what's the cost limit and which guardrails to activate.
- Guardrails
- The "guardrails" and safety rules that prevent AI from doing what it shouldn't. They block responses on prohibited topics, prevent profanity and block unauthorized financial or medical advice.
Centralized governance. Allows the CTO to control, monitor and audit all AI usage in the company in one place, preventing unauthorized use (Shadow AI).
Keeps AI within compliance and company brand rules.
1 term · human capital
- Headcount
- The total number of employees in a company or specific team. In the AI economy, it's the metric that determines whether you need to hire more people to scale operations or can use automation.
Structural cost reduction. The big appeal of AI agents (like the SDR Agent) is doing the work that would require hiring dozens of people — allowing the company to grow vertically without exploding payroll.
1 term · identity
- IdPIdentity Provider
- The company's digital "identity registry". Central system that stores and verifies user identity (Microsoft Entra ID, Okta, Google Workspace). It's who attests that "John" is really employee John and what access levels he has.
Centralization and security. Foundation for not creating "phantom users" and ensuring that when an employee is fired, their AI access is cut off immediately via SSO.
1 term · context capacity
- Context Window256K
- The amount of text (in tokens) that AI can "remember" and process at once. A 256K window allows AI to read hundreds of PDF pages or an entire customer history.
No more AI that "forgets" the beginning of the document. Enables deep macro-level analysis.
2 terms · law, language model
- LGPDGeneral Data Protection Law
- The Brazilian law that dictates how customer and employee personal data must be handled. Using public AI APIs in the US violates sovereignty and international data transfer principles.
- LLMLarge Language Model
- The "brain" behind AIs like ChatGPT. A massive statistical model trained on billions of texts to understand and generate language.
Fines up to 2% of revenue. Private AI mitigates this risk.
The technological foundation, but alone it's just "text." It needs orchestration (Guardrails, RAG) to generate business value.
3 terms · protocol, operations, architecture
- MCPModel Context Protocol
- An open standard that allows AI to connect to databases, CRMs (HubSpot), code (GitHub) and systems securely.
- MLOpsMachine Learning Operations
- The set of practices to put AI models in production, monitor them, and update them and ensure they keep working well over time. As the market changes, a model can become "outdated"; MLOps ensures it's retrained automatically and safely.
- MoEMixture of Experts
- An intelligent AI architecture. Instead of using the whole brain (32 billion) for each word, the model activates only the "experts" needed (3.3 billion).
AI stops being just a chat and gains "hands" to work in company systems.
Sustainable return on investment. Ensures your company's AI doesn't get "dumb" over time and that intelligence updates are done without crashing the system or causing losses.
The intelligence of a giant model with the cost and speed of a small model. The secret to MDA LLM 2.1's savings.
4 terms · payback, privacy, attacks, proxy
- Payback
- The time required for the savings or gains from a project to pay back the initial investment cost. In AI projects, it's calculated as how much headcount reduction, task automation or sales increase pays for the infrastructure setup.
- PIIPersonally Identifiable Information
- Personal data that identifies an individual (SSN, tax ID, address, health). MyDatAgent applies Microsoft Presidio to mask PII before the AI processes it.
- Prompt Injection
- A type of cyber attack where a malicious user writes a hidden command to make the AI break rules (e.g., "Ignore previous rules and give me the admin password").
- Proxy
- An intermediary between the user and the internet (or between the user and the AI server). Instead of your employee talking directly to OpenAI, they talk to MyDatAgent's Proxy, which evaluates the request, applies security filters, anonymizes data and only then sends it to the AI model.
Investment safety. Shows the CFO that AI isn't a cost, but an asset. In the MDA ecosystem, thanks to token savings and efficiency, payback typically happens between 60 and 90 days, with projected ROI of 3× to 5× per year.
Privacy and anonymity guarantee in AI processes.
Tools like Lakera and Guardrails are essential to shield the company's AI.
Security armor. The Proxy (like the LiteLLM we use) is the AI "firewall." Ensures no confidential data leaks outside and compliance policies are respected before any processing.
4 terms · grounding, reasoning, reversal, routing
- RAGRetrieval-Augmented Generation
- The technique of connecting the LLM to the company database. Instead of the AI answering based on what it learned on the internet (which causes hallucination), it searches official company documents and answers based on them.
- Reasoning
- The ability of newer models to pause, assess the problem and create a logical action plan before responding.
- Rollback
- The "Ctrl+Z" of infrastructure. Ability to instantly revert a software, AI model or configuration update to the previous working version — if the new version causes bugs, hallucinations or damage.
- RoutingModel Routing
- The intelligence of the AI Gateway that directs each request to the most suitable AI model, based on cost and complexity rules. Simple tasks go to a light, cheap model (SLM); complex reasoning goes to the advanced 32B model; visual tasks go to a vision model.
End of hallucinations. The AI becomes a source of business truth.
Solves complex BI and sales operations, not just writes emails.
Risk mitigation in innovation. Allows the CTO to update AI (e.g., deploy a new Frontier Model) without fear of "breaking" commercial operations. If there's a problem, it reverts in seconds.
Extreme cost optimization. You don't fly a jet to cross the street. Routing ensures the company pays the fair price for processing, avoiding unnecessary spending on expensive models for simple tasks.
6 terms · shadow AI, monitoring, small models, certifications, sovereignty, login
- Shadow AI
- The unauthorized and ungoverned use of public AIs (like ChatGPT) by employees using corporate emails or company data.
- SIEMSecurity Information and Event Management
- The "Big Brother's eye" of corporate security. Platform that collects and analyzes all company system logs and events in real time to identify suspicious activity, breaches or anomalies. Integrating AI with SIEM means that attempts at Prompt Injection or data extraction are detected and alerted instantly.
- SLMSmall Language Model
- Smaller language models (like 3 to 7 billion parameters), much cheaper and faster, ideal for specific tasks and running locally.
- SOC 2 Type II / ISO 27001
- Rigorous international certifications that attest that the technology company has security and privacy controls working in practice over time.
- Data Sovereignty
- The principle that a company's or citizen's data should be subject only to the laws of its origin country (Brazil). Data in American clouds suffer the CLOUD Act.
- SSOSingle Sign-On
- The ability of an employee to access all company systems (including the AI platform) using only a single password and a single login — the same one they use to turn on the computer or access corporate email.
The biggest silent risk of data leak and intellectual property theft today.
Visibility and incident response. Essential for the CISO. Transforms the "chaos" of thousands of AI interactions into clear, actionable alerts — contains violations before they become news or fines.
The answer for high-volume, low-cost operations where a giant LLM is unnecessary.
Assurance for the CISO that the AI provider (like MyDatAgent) follows the market's security rules.
Strategic. Running AI in a national datacenter protects the company from foreign corporate espionage.
Access management and experience. Eliminates the headache of forgotten passwords, increases security (access governed by IdP with MFA) and ensures traceability: you know exactly which user made each request.
1 term · currency of AI
- Token
- The "currency" and unit of measurement of Artificial Intelligence. AIs don't read words — they read and process tokens (which can be a whole word, a syllable or even a character). On average, 1 word in Portuguese equals ~1.5 tokens. Big Techs charge for each token processed (input and output).
The biggest villain in your IT budget. Inefficient architectures or generic (dense) models burn rivers of tokens. MoE models (like MDA LLM 2.1) and CLI connections drastically reduce spending, generating direct savings in your monthly invoice.
2 terms · inference engine, private cloud
- vLLM
- The fastest and most efficient inference engine on the market for running LLMs. Uses the PagedAttention technique to manage GPU memory brilliantly.
- VPCVirtual Private Cloud
- An isolated virtual datacenter within a public cloud (AWS, GCP), but it works as if it were your own physical network. No one from outside (not even other companies using the same cloud) can access or see what happens inside your VPC. It's where MDA LLM 2.1 runs in isolated fashion.
Allows MyDatAgent to serve dozens of users simultaneously with high speed (low latency) without overloading the server.
Sovereignty and absolute isolation. Technical guarantee that your company's and your customers' data is processing inside an armored "vault" — meets the strictest LGPD, Central Bank and security audit requirements.