VeloDev
AI Engineering & Integration

AI engineering for software agencies and SaaS teams. I add agents and AI features to products that already exist, either as your technical partner or as a white-label delivery partner.

VeloDev designs and builds custom AI agents and AI-powered applications that integrate with your existing software and workflows.

I reply within one business day.

/ capabilities

Three kinds of work, all built on the same engineering basics: authentication, permissions, logging and tests.

Custom AI Agents

Agents with controlled access to your tools, APIs and data. Every permission is explicit, and sensitive actions wait for human approval.

Where a process really has distinct stages, I can split the work across specialised agents. When one agent is enough, I’ll say so.

AI-Powered Applications

AI features built into web, mobile and SaaS products: search, assistants, extraction, summaries, with proper auth and per-user permissions.

AI Workflow Automation

Integration with the systems you already run: documents, CRM / ERP, inboxes and internal processes, with logging and fallbacks.

/ use cases

Illustrative solutions I can build. These are examples of approach, not delivered client projects. The built products are at the top of the page; other real work is linked under “Proof” below.

Example solution

Document Processing Assistant

Extracts, verifies and structures data from incoming documents. A person reviews and corrects before anything is accepted.

Example solution

SaaS Knowledge Assistant

Retrieval-augmented answers over your documentation, respecting each user's permissions, with sources shown.

Example solution

Business Operations Agent

Reads from your APIs and proposes actions. Sensitive operations require approval before they run.

Example solution

AI Feature Integration

Adds an AI capability to a product you already have, without rebuilding it.

/ process

Seven steps in four phases, from first conversation to a system running in production.

  1. Phase 1

    Assess

    A few days

    Understand the workflow and decide what is worth building.

    • 01
      Discovery

      The workflow, the users and what “better” would mean.

    • 02
      Technical Assessment

      Data, systems, risks, and whether AI is the right tool.

    • 03
      Architecture

      The simplest design that meets the requirements.

  2. Phase 2

    Prototype

    1–3 weeks

    A first working slice on one real workflow.

    • 04
      MVP

      A working slice on real workflows, built in small increments.

  3. Phase 3

    Build & evaluate

    3–6 weeks

    Grow the working slice into the full build, then measure it.

    • 05
      Testing & Evaluation

      Test sets, edge cases and measured output quality.

      Outcome: an evaluation set and a results report.

  4. Phase 4

    Run

    Launch + optional monthly plan

    Go live with monitoring, then keep it healthy.

    • 06
      Deployment

      Release with monitoring, logging and a rollback plan.

    • 07
      Maintenance

      30 days of fixes included, then an optional monthly plan for model and cost review and improvements.

Typical ranges; they depend on scope.

Optional paid technical assessment

For complex projects you can start with a paid assessment that takes a few working days and ends in defined deliverables. You keep all of them whether or not you build with me, so you can also take the plan to another team.

I may recommend not building something. If the honest answer is that AI isn’t the right tool, the assessment says so.

  • Written architecture proposal
  • Risk and data-access review
  • Scoped MVP plan with milestones
  • Effort estimate for the build
/ ownership

Clear terms up front: what is yours when the project ends, and what sits outside my fee.

You own the work

  • The code
  • The prompts
  • The evaluation sets
  • The deployment configuration

No licence fee, no lock-in.

Billed by the provider

Model API usage is billed directly by the provider to you. It is not part of my fee, so that cost stays visible to you.

After launch

30 days of fixes are included. After that, an optional monthly plan covers fixes, model and cost reviews and improvements, scoped separately.

/ for agencies

Need AI delivery capacity behind your own brand? I work as a white-label partner for software agencies, as one senior engineer you can rely on.

  • You keep the client and set the price.
  • I build behind your brand.
  • No contact with your client unless you approve it.
  • No branding of mine in code or deliverables.
  • NDA and data-processing agreement available on request.
  • Handover documentation and a runbook, so the project doesn’t depend on me being in the room.

A paid pilot is the usual first step: one scoped piece of work, so you can see how it runs before anything bigger.

Talk about a partnershipI reply within one business day.
/ architecture

A typical shape for an agent system that touches real business data. Not every project needs multi-agent: I choose the simplest solution that meets the requirements, and often that is a single agent or a plain AI feature.

The ladder: climb only as far as the problem needs
  1. Direct model call
    One prompt in, one answer out
    Executed by: model
  2. Workflow
    Fixed steps in code, model calls inside
    Executed by: code
  3. Default starting point
    Single agent with tools
    The model chooses among tools you permit
    Executed by: agent
  4. Multi-agent
    Only when justified
    Executed by: agent
Reference flow, at the multi-agent end of the ladder
User / Business App
Executed by: person
Authenticated API / AI Gateway
Executed by: code
Agent / Orchestrator
Executed by: agent
Research Agent
Executed by: agent
Data Agent
Executed by: agent
Workflow Agent
Executed by: agent
Tools, APIs, Data
Executed by: code
Validation & Logs
Executed by: code
Human Approval (when needed)
Executed by: person
Final Result
Executed by: code
Across every step
  • Guardrails
  • Eval set
  • Traces
  • Cost
Requests are authenticated, agents only reach the tools they need, output is validated and logged, and a person approves sensitive steps. Tags show who executes each step: plain code, a single model call, an agent that loops over tools, or a person. In the single-agent default, the orchestrator and the specialists collapse into one agent.
/ why velodev

VeloDev is led directly by Marian Putirac, a senior full-stack engineer (React / Next.js, Node, headless WordPress) working from Romania. You get one accountable person, not a hand-off chain.

StackNext.js / TypeScript, Node.js, Supabase / PostgreSQL, Anthropic Claude, OpenAI and Gemini APIs.

Senior full-stack ownership

Projects are led directly by one senior full-stack engineer, end to end: frontend, backend, integration and deployment.

Security by design

Authentication, permissions, data protection and secret management are part of the architecture, not an afterthought.

Reliability you can inspect

Evaluation sets, logging, monitoring, fallbacks and cost tracking, so behaviour and spend are visible.

Direct technical communication

You talk to the person who writes the code. Clear ownership, no layers in between.

Proof: AI work in live products
  • Investment Analysis Tools

    Full-stack stock analysis platform I built solo, with Gemini AI recommendations.

    Case study →
  • Mechanic Copilot

    AI assistant for car workshops that suggests an investigation plan and leaves the diagnosis to the mechanic. Gemini with structured output; I built it end to end.

    Case study →
  • Incom AI

    B2B platform using the OpenAI API to automate customer communications and feedback analysis. My role: technical contributor (Next.js / React features, OpenAI API integration).

    incom.ai ↗
/ faq

Can you integrate AI into an existing application?

Yes. AI capabilities can usually be added through your current APIs and auth, without a rewrite. The assessment step checks what the codebase and data can support before anything is built.

When is a multi-agent architecture useful?

When a process has distinct stages that need different tools, data or permissions, and a single agent becomes hard to control or test. Not every project needs multi-agent. I choose the simplest design that meets the requirements.

How do you handle wrong or invented answers?

I can’t promise a model never gets something wrong, and I don’t offer guaranteed accuracy. What I do: agree an evaluation set with you up front, validate outputs before they are used, require human approval for sensitive actions, and log runs so errors can be traced. Quality is measured on a test set I agree with you in advance, and you get the results.

Is my data used to train models?

It depends on the model provider and the API tier, and provider terms change, so I don’t make a blanket promise. For each project I agree with you which data goes to which provider, go through that provider’s current terms for the tier in use, and put the outcome in writing.

Who owns the code and prompts?

You do. The code, prompts, evaluation sets and deployment configuration are yours, with no licence fee and no lock-in. Model API usage is billed by the provider directly to you and is not part of my fee.

How long does a first project take?

Typical ranges: the assessment takes a few days, a first prototype 1–3 weeks, and the build with evaluation 3–6 weeks. They depend on scope, integrations and the state of the data. I give a concrete estimate after the assessment.

What affects the implementation cost?

Scope, number and quality of integrations, state of the data, security requirements, how rigorous the evaluation needs to be, and ongoing model usage. A scoped estimate follows the assessment, not before.

How do you protect business data?

Access is limited to what the task needs, secrets stay server-side, permissions follow your existing user model, and sensitive actions can require human approval. I agree with you up front which data goes to which model provider.

What happens after deployment?

The system ships with logging and monitoring. 30 days of fixes are included. Afterwards I can stay on through an optional monthly plan for fixes, model and cost reviews and improvements, scoped separately.

Can a project start with a smaller proof of concept?

Yes, and I usually recommend it. A narrow proof of concept on one real workflow shows quickly whether the approach is worth building out.

I reply within one business day.