Custom AI & Python builds
A senior AI engineer for hire, not an agency
Quick answer: I build production AI systems for founder-led businesses. Retrieval over your own documents, grounded assistants that cite their sources, document intelligence, and automation that snaps onto the tools you already run. You work directly with the person who writes the code, and you can start with a small paid pilot instead of a leap of faith.
What I build
Four kinds of work, one discipline underneath: the AI only speaks from material you control, and every claim is checkable.
RAG & semantic search over your documents
Ask questions across your contracts, reports, filings, or knowledge base and get answers grounded in your files, not the open internet. Vector search (pgvector, embeddings) finds the right passage by meaning, not keyword, and every answer is verified back to its source page.
Grounded LLM & agent apps
Assistants and agents that cite their sources and refuse to guess. Anti-hallucination is the architecture, not a footnote.
Document intelligence
OCR plus structured extraction that turns scans, PDFs, and messy forms into clean, validated, queryable data.
Automation & integrations
The AI layer that snaps onto your existing CRM, storage, and internal tools. No rip-and-replace.
Private by default
Your documents do not have to leave your infrastructure. I run open models locally for the parts that touch sensitive material, so confidential deal, patient, or regulatory files are never sent to a third-party AI provider. Where a frontier model genuinely earns its place, you decide what reaches it.
Costs that cannot run away
Every build ships with per-run cost caps, usage metering, and alerting, because an AI system that quietly triples its own bill is not production software. You see what each run costs before you scale it.
How it works
The honest path is proof, not a contract. You start small, judge a working result on your own data, then decide on the larger build.
- Async intake. You send one workflow that hurts through a short written form. No call.
- Paid proof pilot. One clear deliverable, fixed price from $2,500, about one to two weeks, on your real data.
- Full build or fractional. If it works, we extend it into a fixed-scope build, or I stay on as a fractional AI engineer. The pilot fee is credited toward the build.
Investment
Legible pricing, not a discovery-call mystery. Three ways to work, and you always know the number before we start.
Paid proof pilot
from $2,500
- •One clear deliverable
- •Fixed scope, ~1 to 2 weeks
- •Run on your real data
- •Fee credited toward the full build
Fixed-scope build
$8,000 to $25,000
- •One defined system
- •Clear spec + acceptance test
- •Time-boxed delivery
- •You know the number up front
Fractional AI engineer
from $4,000/mo
- •Or a day rate from $750
- •Ongoing build capacity
- •On call for your AI roadmap
- •Scale hours up or down
All prices are in US dollars and exclude VAT. Romanian clients have Romanian VAT added on top. EU businesses outside Romania with a valid VAT number are normally handled under the reverse charge, and clients outside the EU pay no EU VAT. Invoiced by DEVALAND MARKETING S.R.L., VAT number RO50841395. Your exact treatment is confirmed in writing before anything is signed.
Proof, not promises
These are real, shipped systems, not demos. The point of hiring at this level is that the exact kind of system you need has already been built and is running in production.
Deal OS
A cited M&A diligence platform that reads 100-page deal documents, runs OCR and extraction, and returns findings with sources attached under a verify-not-summarize rule. See Deal OS.
Amy
A grounded product-Q&A voice assistant running live for a Shopify brand, quoting real prices from a live catalog sync instead of inventing a number that sounds right. Hear a live voice AI.
Nadia & Delia, website concierge AIs
Source-gated chat assistants that answer visitor questions from a real knowledge base and refuse to invent policy, escalating to a human when they cannot answer. Nadia runs on Devaland's own site (she is the assistant in the corner of this page) and answers in seven languages; Delia runs the Goldlett store. See our AI agents.
Who this is for
The work fits best with bootstrapped founder-operators, often running a multi-entity group or holding, with heavy document, regulatory, or data workflows and a real budget. Biotech and diagnostics, professional services, roll-ups, and M&A-active acquirers are typical. If a wrong answer in your business costs real money, that is exactly the kind of problem these builds are for.
Frequently asked questions
What do you actually build?
Production AI systems for founder-led businesses: retrieval (RAG) and semantic search over your own documents using vector search and embeddings, LLM and agent apps with anti-hallucination, document intelligence (OCR plus structured extraction), and workflow automation that snaps onto the systems you already run. Not slideware, working software you can judge on your own data.
Are you an agency?
No. You work directly with one senior AI and Python engineer who writes the code and is accountable for the result. No account managers, no juniors learning on your budget.
How do we start without a big commitment?
With a paid proof pilot: one clear deliverable, fixed price from $2,500, about one to two weeks, run on your real data. If you continue, the pilot fee is credited toward the full build. You buy proof before you buy a project.
What does it cost?
Paid proof pilot from $2,500 (credited to the build). Fixed-scope build typically $8,000 to $25,000 depending on scope. Ongoing work as a fractional AI engineer from $4,000/month or a day rate from $750. You know the number before we start. All prices are in US dollars and exclude VAT: Romanian clients have Romanian VAT added, EU businesses outside Romania with a valid VAT number are normally handled under the reverse charge, and clients outside the EU pay no EU VAT.
Do my documents get sent to OpenAI or Anthropic?
Only if you want them to. I run open models locally on infrastructure I control, so the parts of a pipeline that touch confidential material can be handled without your files ever leaving that environment. My own platform embeds and indexes documents on a local model for exactly this reason. Where a frontier model is genuinely the right tool, we agree in advance what reaches it and what never does, and that boundary is written into the build.
What stops the AI bill from running away?
Per-run cost caps, usage metering, and alerting, built in from the start rather than bolted on. Every run has a hard ceiling, spend is visible per workspace, and I get paged before you get surprised. I run this on my own production systems, so the cost discipline is not theory.
How do you stop the AI from making things up?
Architecture, not hope. Every claim is grounded in your source material and cited back to where it came from, with human approval for anything consequential and evaluations that measure accuracy. The rule on every build is simple: cite the source or cut the claim.
Do I have to get on a call?
No. Intake is async through a short written form. Marius reads it, tells you honestly whether a build makes sense, and if it is a fit you scope one clear deliverable in writing. No sales calls.
Who is this for?
Two kinds of client. First, bootstrapped founder-operators, often running a multi-entity group or holding, with heavy document, regulatory, or data workflows and a real budget. Biotech and diagnostics, professional services, roll-ups, and M&A-active acquirers are typical. Second, agencies, consultancies and data teams who need an AI delivery arm: I build under your brand, in writing only, and the client relationship stays yours.
Do you work white-label for agencies?
Yes. Agencies, consultancies and data teams bring me in as the AI build partner behind their own delivery. You keep the client and the brand, I build the system. I can work inside your client thread under your name, or stay entirely invisible and deal only with you, whichever you prefer. Everything is in writing, and the same paid pilot terms apply so you can quote your client a fixed number before committing.