AI agents and automations built by DeviDevs

DeviDevs builds AI agents for companies: agents that split the work and check each other, assistants that answer from your company's documents with the source in view, and automations across email, e-invoicing, ERP and CRM, including MuleSoft, Salesforce and SAP. Our own company runs every day on a multi-agent system we built on Claude.

Talk to us about an agent for your company

DeviDevs, internal system

A manager agent splits the request between specialised agents (code, research, copy). An independent reviewer reads their work before anything leaves the company.

Challenge
A do-everything agent makes mistakes without noticing, and nothing stops it.
Outcome
Our inbox and code run through it daily, and our invoices whenever we issue them. We build the same pattern on your company's processes.
Read the case study
  • Claude
  • Claude Code
  • MCP
  • Python

Multi-agent AI systems: agents that check each other

Agents that take one part of the daily work and hand it on, to a person or to another agent.

who does whatthe requestmanagercodetextresearchverifierhuman: yes / noout

  1. DeviDevs, internal

    Jira tickets get worked every day

    Challenge
    Small tickets pile up, and each needs someone to work it.
    Outcome
    The agent picks the tickets it can finish, does the work (code, documents, checks) and leaves it on the ticket. Emails and invoices it prepares for approval.
    • Claude Code
    • Jira
    • Git

  2. DeviDevs, internal

    Incoming email arrives sorted by urgency

    Challenge
    In a full inbox, a client's request gets lost among notifications and ads.
    Outcome
    Several times a day, new emails are read, sorted and turned into tasks, with replies first.
    • Gmail
    • Jira
    • Claude

  3. DeviDevs, internal

    Decisions that need a person reach the phone

    Challenge
    Some decisions are not the agent's to make, and the person is not always at the desk.
    Outcome
    The agent asks on Telegram when it needs a yes. Voice replies are transcribed locally and handled like written requests.
    • Telegram
    • Claude

RAG: assistants that answer from company documents

Our agents search what the company knows before they act. We build the same on your contracts, procedures and archive.

memory (qmd)the question?keywordsmeaningnotes · decisions · emailanswer + source“not found” only after searching everywhere

  1. DeviDevs, internal

    Agents answer from the company's notes, with the source in view

    Challenge
    An agent without memory repeats mistakes already solved.
    Outcome
    Before deciding, agents search the company's notes, rules and decisions, by keyword and by meaning at once, and answer with the source.
    Read the case study
    • RAG
    • Hybrid search
    • Embeddings
    • MCP

  2. DeviDevs, internal

    A "we don't have that" only comes after every source was searched

    Challenge
    Information sits in email, documents, Drive and tickets; one source alone gives a wrong answer.
    Outcome
    One search covers every source and reports which one did not respond, so an empty result is never taken as proof.
    • Hybrid search
    • Gmail
    • Google Drive
    • Jira

  3. DeviDevs, internal

    An error fixed once is never debugged from scratch again

    Challenge
    The same error comes back, and whoever fixed it has forgotten how.
    Outcome
    Every proven fix goes into a register that agents read before debugging.
    • RAG
    • Claude Code

Automation and integration: e-invoicing and the tax authority

Flows that connect systems, so data stops being copied by hand.

the invoice goes by itselfinvoiceXML checkedANAF · SPVreceiptnew pagesearch engines

  1. DeviDevs, internal

    Our invoices reach e-Factura through the API, not the SPV portal

    Challenge
    Every invoice must reach the national system, as strict XML, within five working days.
    Outcome
    Invoices go out through the API after a person confirms, as UBL 2.1 under the CIUS-RO rules, and get an upload index from ANAF. When ANAF does not respond, the invoice waits in a queue and is not lost.
    Read the case study
    • UBL 2.1
    • CIUS-RO
    • OAuth2
    • Python

  2. DeviDevs, our own sites

    New pages are sent to search engines on release

    Challenge
    A new site can wait a long time before a search engine sees it.
    Outcome
    On release, pages go to Bing and the engines that use IndexNow. Google Search Console shows which pages do not rank, and for which searches.
    • Google Search Console
    • IndexNow
    • Next.js

AI agent guardrails and human approval

Rules that stop an agent in time, and measurements of what it does.

the guardsthe agentsafe? goesmoney · contractsa human decidescost / run

  1. DeviDevs, internal

    Agents work alone on what is safe and stop for a person where it matters

    Challenge
    An agent that sends emails or issues invoices has to be stopped before the mistake.
    Outcome
    Rules in code, with their own tests, check every agent action before it happens. Payments and contracts wait for a person's yes.
    Read the case study
    • Guardrails
    • Automated tests

  2. DeviDevs, internal

    We know what every run costs and how often a person corrects the agents

    Challenge
    An agent that seems to work can cost more than the work it takes over.
    Outcome
    Each run's cost is recorded, usage has alert thresholds, and the correction rate is measured weekly from real conversations.
    • Evals
    • Python

Web platforms, websites and online stores

Sites and shops we made, plus the platforms we use in the company.

a site from usresearchmockupsverifierlive

  1. DeviDevs, how we work

    How we build a site: research, mockups, an independent reviewer

    Challenge
    A template site looks like all the others, and mistakes show up only after launch.
    Outcome
    We start from what the company's customers search for and the best sites in the field, pick the direction on mockups before code, and run copy, speed and SEO checks. It goes live only on the owner's yes.
    Read the case study
    • Next.js
    • Tailwind
    • Vercel

  2. Bio cosmetics and supplements, Romania

    Farmacist pe bune: online shop with an editorial section

    Challenge
    Orders and articles with no monthly platform fee, for a non-technical owner.
    Outcome
    Live shop with a product catalogue and an editorial section.
    Read the case study
    • Static HTML
    • Supabase
    • Stripe
    • Vercel

  3. Martial arts club, Bucharest

    Ki Aikido Sakura: a web address instead of a Facebook link

    Challenge
    Everything lived in Facebook posts, where a newcomer could not find the schedule or the location.
    Outcome
    The schedule, the location and the trial-class request are in one place, indexed by search engines.
    Read the case study
    • Next.js
    • Tailwind
    • Vercel

  4. DeviDevs AI Agents Lab

    The certificates we issue can be verified online by code

    Challenge
    A PDF certificate can be forged, and whoever receives it has no way to check it.
    Outcome
    Each certificate has an ID and a QR code that lead to its page on devidevs.com, with the name, ID and details confirmed by DeviDevs.
    • Next.js
    • QR code

Education: AI Agents Lab and student internships

What we use in the company we also teach to students.

DeviDevs AI Agents Lab

Certificate of participation

First Last

built and shipped an app with AI agents

verification codeDD-LAB-••••sample

  1. DeviDevs educational program

    AI Agents Lab: from the first coding agent to a published app

    Challenge
    People who want to learn AI agents find scattered tutorials and rarely a full program.
    Outcome
    Participants worked with technical mentor Nicu Constantin on each module's tools and on practical projects. Their certificate carries their name and is verifiable online.
    • Claude Code
    • RAG
    • Playwright

  2. Education, Bucharest

    UTCB students did their internships with us

    Challenge
    The students needed real projects and someone following their work.
    Outcome
    They worked with tutors from our team, and UTCB lists DeviDevs among its career center partners.
    • Internship
    • Mentoring

What we build for companies

From the first agent to the whole platform.

Custom AI agents
A specific job in your company, wired into the systems you already use.
Multi-agent systems
Agents that split the work and check each other.
RAG on your company's documents
Assistants that answer from your contracts, procedures and archive, with the source.
Automations between systems
Email, Drive, Jira, ERP and CRM connected, with no copying by hand.
e-Invoicing and ANAF integration
Invoices sent automatically into Romania's national system, with a queue when ANAF does not respond.
ERP and CRM integration
Agents that read from and write to your ERP and CRM through their APIs.
Enterprise integration
APIs and flows on MuleSoft (certified Level 2 developer), Salesforce and SAP.
Web and internal platforms
Apps, portals and internal tools, from prototype to production.
MLOps and model serving
Training, model registry, automated delivery with rollback, monitoring.
Computer vision
Image classification and detection, real time or in batches.
LLMs from several providers
Claude, OpenAI and Google in one system, with a fallback when one is down.
Evals and cost per agent
Tests on what agents produce and the cost of every run, from day one.

AI security and the EU AI Act: how we work

On request, on your systems, so they are not in the register above.

AI security assessment

The first assessment is passive, with no active pentest: DNS, the web app, email security, exposed code and data, the AI threat model and your EU AI Act position. You get the report and the fixes, in priority order.

See the service

EU AI Act readiness

Risk classification (Annex III and Article 50 transparency), gap analysis, documentation and the audit trail, until your system can be audited.

See the service

AI agents: frequently asked questions

What is an AI agent and what does it do?

A program connected to your company's systems that reads, decides on written rules and moves the work on. Payments and contracts wait for a person's yes.

What is the difference between an AI agent and a chatbot?

A chatbot answers questions. An agent finishes a job inside your systems, from reading the email to filling in the document, and an independent reviewer checks its work before it goes out.

How much does an AI agent cost?

The price is fixed, per package, and published on our pricing page, from scanning one process up to a production agent and maintenance.

How long until it works?

It depends on the package. Each one has its timeline next to the price on the pricing page, agreed before we start.

Does it connect to the software we already use?

Yes. Our agents already work with Romania's e-invoicing system, Gmail, Google Drive and Jira. For ERP and CRM we connect through their APIs, including MuleSoft (certified Level 2 developer), Salesforce and SAP.

Is our company data safe?

Rules in code check every agent action before it happens, and payments and contracts wait for a person's yes. On request we also run an AI security assessment.

What happens if the agent makes a mistake?

An independent reviewer reads its work before it goes out, and the agent stops for a person where a mistake costs. Fixed errors go into a register read before debugging.

What exactly do you do: audit, documentation or implementation?

All three, in steps: scanning one process, the diagnosis, the build and maintenance. We also run AI security assessments and EU AI Act readiness.

Have a process worth handing to an agent?

Tell us what your team does by hand today. We will come back with what can be taken over and the next step.

Tell us about your process

Updated