DMDejan Milošević
Hire me
Open for contract work · 48h pilot · Belgrade (CET)

I design and build AI systems that make businesses run better.

From business process analysis to production-ready automation. I find the bottleneck, design the system, direct the agents that build it, and stand behind what I control. Agents write the code. I own the outcome — the part we wrote into the contract.

26
tests green
200 rps
p95 16 ms
50
incidents in memory
5/5
e2e runs green

Agents write the code.
I own the outcome.

Ownership here is concrete, not a slogan. I sign for what I control and what we write down: the design, the tests, the delivery, and how the flow behaves in production. Spec Machine is my internal exoskeleton — a machine that remembers every failure, with a human who signs off. I do not promise your revenue or your market, and I do not hide who typed the code.

01 — what I take off your hands

Six problems I can take off your hands.

You do not need to know whether this is AI, automation or custom software. You need the problem to stop.

01

Manual business processes

Employees copy data between email, spreadsheets, forms and business software — every day, by hand.

02

Slow document processing

Invoices, orders, reports and forms need repetitive manual review before anything can move.

03

Disconnected systems

CRM, ERP, accounting software and internal tools do not talk to each other, so a person is the integration.

04

Uncontrolled AI usage

The company experiments with AI, but there are no reliable workflows, permissions or audit trail.

05

Operational blind spots

Management has no timely view of delays, errors, costs and performance — answers arrive too late to act.

06

Custom internal tools

Off-the-shelf software does not fit how the business actually works, and nobody is building the missing piece.

Not sure which one is yours? That is where we start. Write one paragraph about the process and I will tell you what is worth automating — and what is not.

01 — why it pays off

Five pillars, each with a number behind it.

A machine alone cannot promise anything. A person with a system can promise, prove and sign.

01
Ambition

I don't ship it working. I ship it surviving.

A junior delivers something that runs on their laptop. I deliver FOR UPDATE SKIP LOCKED, a dead-letter queue, retries with jitter and a test for every path. That is why the pilot can carry a money-back guarantee at all.

26 tests · DLQ replay · p95 16 ms
02
Imagination

I see a solution where others see paperwork.

You see 200 invoices a month. I see PDF → OCR → classification → database → report, with no human in the loop. That is why Company Intelligence maps the process first and picks technology second.

invoice OCR in daily use
03
Restraint

I don't spend your budget on trends.

No Kubernetes for three webhooks. Postgres, a rule book and a Docker sandbox on a VPS do the job. Only the Saboteur gets an expensive model, because that is the one that has to break bad architecture. Result: about $0.004 per run, not $0.15.

≈ $0.004 / run
04
Logic

Every decision has to survive the Saboteur.

There is no “the AI said so”. Every design passes at least three ADRs, plus idempotency, quiet hours, opt-out and 50 recorded incidents. If it does not pass, the job is needs_human — not done. That is why you get no duplicates and no flaky automation.

50 incidents · needs_human fallback
05
Craft

From the conversation to production, alone.

Business Analyzer, Solution Architect, Builder that writes SQL, zod contracts and the workflow, tested in an isolated container, and a Scoper that hands you the quote with dates. You talk to one person and get the whole chain.

one person · whole chain
the short version

Spec Machine is not an autopilot — it is an exoskeleton. I find the company and understand the process; the machine maps it and finds where time leaks. I ask the questions; the machine searches 50 incidents so we do not repeat the mistake from 22 October. I make the call; the Saboteur is there to break it if it is weak. I sign; there is an approve button, never an auto-deploy.

02 — what I do

Three things, done properly.

01

Intake & qualification

Webhook that answers in 300 ms, idempotency that actually works, DLQ with replay. The lead that arrives twice is not a second lead.

02

Agent pipelines

Architect, historian, saboteur, builder, scoper. Every decision cites the rule and the incident that paid for it — not “best practice”.

03

Ops you can show

Trace per job, cost per stage, retry counters, dead letters. When it breaks at 3 a.m. there is a log, not a guess.

03 — selected work

Three systems, with the evidence attached.

What the problem was, what I built, what my role was, and how you can check it yourself. Anything not measured is not claimed.

In production · used daily

MAJABRAJAN — invoice OCR

Problem
Incoming invoices were read and typed into the accounting system by hand, with transcription errors.
Solution
OCR → field extraction → classification → export in the accounting software's format, with a review queue for anything uncertain.
My role
Architecture, development, integration with the accounting workflow.
Proof
Video of the real process, a sample invoice and the extracted data.
Result
Volume processed and recognition rate are measured inside the tool; no public number is claimed here.
Running · 26 tests green

AI Lead Intake — n8n + 4 agents

Problem
Webhook retries sent the same lead twice, SMS went out at 22:00, and nobody could prove what happened.
Solution
Idempotent intake (sha256 of source + event id, never a timestamp), Postgres outbox, retry with backoff, DLQ, opt-out and quiet hours, Telegram human approval before any message leaves.
My role
Architecture, schema, workers, compliance rules, tests, deployment.
Proof
7-node flow, Postgres rows with idempotency key and trace id, a 20s recording of a real run.
Result
No duplicate lead in 26 tests, no message outside quiet hours, p95 16 ms at 200 rps.
Used weekly · spec engine

Spec Machine — problem to signed spec

Problem
Design decisions rested on one person's memory; the same failure got rediscovered every few months.
Solution
Five stages — architect, historian, saboteur, builder, scoper. Every ADR must cite a rule (R01–R33) and an incident from the 50 recorded ones, or the job stops at needs_human.
My role
Design of the swarm, the rule book, the incident memory, the approval loop.
Proof
Trace per job with model, tokens and cost; a 15s recording of a run ending in [R01 + C1].
Result
≈ $0.004 per run, no design decision ships without a citation.
04 — how it is proven

You can open all of it.

The command center is not a mockup: it runs on the same database as the swarm, and every number below is read from it.

livespec-swarm

Rules with a price tag

Every ADR cites the rule and the incident that paid for it — [R01 + C1]. If the model ignores history, the job stops and waits for a human instead of inventing advice.

  • [R01 + C1]idempotency key = sha256(payload) — the duplicate webhook cost two hours once
  • [R17 + F6]quiet hours in the recipient's timezone, never the server's
  • [R09 + B1]4xx means zero retries: straight to dead letters with an alert
50 incidents · needs_human fallback
26 testsintake

Duplicate sends one SMS

Idempotent webhook, outbox with 1s/5s/25s/125s backoff, dead letters with replay.

200 rps · p95 16 ms
5/5e2e

Verified five times

13 modules, 6 interactions, zero console errors — on a fresh database, from scratch.

clean-room run
05 — how it goes

Four steps. No calls.

01

Send one flow

The thing your team still does by hand. One paragraph.

02

Scope, in writing

Input, output, success criteria, price, start time. Same day.

03

48 hours

Progress notes instead of questions. I ask once, in writing.

04

Delivery

Code, schema, tests, a 2-minute Loom, 14 days of fixes.

free · no account · 2 minutes

Does your AI system need paperwork under the EU AI Act?

Twelve questions. You get the risk level, the articles that apply (5, Annex III, 50) and the controls you are still missing. It is the same classification I run before quoting an evidence pack.

Run the free check ↗
05 — how I work

From a business problem to a working system.

Async-first by default: written updates, no status meetings. Calls and workshops when the project actually needs them — I do not avoid people, I avoid unnecessary meetings.

01

Understand

We look at the process as it runs today: tools, people, bottlenecks and what the business actually needs.

02

Design

You get a written solution architecture: scope, timeline, acceptance criteria and a cost estimate before anything is built.

03

Build

I develop, integrate and test against those acceptance criteria — with tests, logs and a recording of it running.

04

Operate & improve

Monitoring, documentation, maintenance and the next improvement, in a monthly plan.

06 — why work with me

What you are actually buying.

Business-first thinking

I understand the process before I pick the technology. If the answer is a form and a rule, I say that instead of selling an agent.

End-to-end ownership

Analysis, architecture, development, integration, delivery. One person responsible for the whole chain, not a handover between three vendors.

Built for real operations

Tests, logs, error handling, security and documentation are part of the delivery — not a later phase nobody buys.

Direct collaboration

You talk to the person who designs and builds the system. No account manager, no translation layer.

07 — engagement models

Three ways to work together.

Start with an assessment if you are not sure what to build, a pilot if you are, or a partnership if the systems keep growing. Prices below are proposals — the real number follows the scope.

Entry

Process Assessment

$249
fixed scope, no further obligation

For a company that knows something is wrong but not yet what the right solution is.

  • One business process analysed end to end
  • Bottlenecks and cost of the current way identified
  • Proposed AI / automation architecture
  • Cost estimate and realistic savings
  • Written implementation plan you can keep
Start with this
Ongoing

AI Systems Partner

by agreement
monthly engagement

For companies that need someone to own the AI systems long term.

  • New automations, built and maintained
  • Integrations and improvements to what exists
  • Monitoring and incident handling
  • Technical documentation and quality control
  • Monthly plan and priorities
Start with this

Payment by card through Creem, or a contract with an invoice — same price, different paperwork. Company details, purchase orders and NDA on request.

07 — questions

The four that get asked.

Why 48 hours?

Because a flow I cannot harden in two days is a flow I should not take as a pilot. Bigger work is scoped after.

What if it does not work?

The outcome is written before I start. If I miss it, the money goes back and the code stays with you.

Do you only do AI?

I do the plumbing AI needs to survive: queues, retries, idempotency, compliance, audit. The model is the easy part.

Can you join full-time?

Yes — that is the point. Contractor first, seat second. Remote, CET, 4h overlap with EST.

Send one flow you hate
doing by hand.

I reply with a fixed scope and a start time — same day, in writing.

You keep the code No hourly billing No meetings Money back on the pilot
08 — about

Independent AI systems architect.

I work at the intersection of business operations, software architecture and AI automation. My focus is not another chatbot — it is systems that solve an operational problem and can be tested, maintained and used in production.

Before software, I spent years in organisations where process, measurement and standards were the job: laboratories, quality systems, international certifications. That is where I learned to ask what "it works" actually means, and to write it down before building anything.

LocationBelgrade, Serbia (CET)
CollaborationRemote, international
EngagementProject / contract / ongoing partnership
LanguagesSerbian, English
AvailabilityOpen for new projects
who I work with

Built for companies with real operational complexity.

Businesses that have outgrown spreadsheets, manual processes and disconnected software — and need systems that fit how they actually operate.

  • Companies with 10–200 employees
  • Accounting and administrative firms
  • Trade, distribution and logistics
  • Manufacturing and technical organisations
  • Companies using several systems with manual work between them

Not sure whether your process can be automated? That is where we start.

09 — project inquiry

Let's understand your business problem.

Tell me what isn't working, how you handle it today and what you want to improve. I review the request and come back with a recommended next step — not a generic sales call.

01 — your details & need02 — process, impact, expectations
What kind of help are you looking for?
What is the main challenge?
Two short steps. Nothing is sent until the end.