Selling complex technology
Pre-sales and solution design for vendors and integrators: positioning, the technical narrative, and a business case that survives the CFO's second question.
Independent technology consultant · Serbia & Spain
Most technology fails in the meeting, not in production.
Independent consulting on cloud, machine learning and applied AI. The solution gets designed, then explained to the people who have to sign for it — in plain language, with the economics worked out.
The work
Two languages, one decision
Engineering describes a system. The business decides on an outcome. Most projects stall in the gap between those two sentences — and that gap is where I work. Every proposal exists in both columns at once.
Managed Kubernetes, autoscaled across availability zones
A traffic spike stops costing you a night of downtime — and you stop paying for idle capacity the rest of the year
Speech-to-text and synthesis pipeline, tuned on your own domain vocabulary
Every customer call becomes searchable text, and your agents stop writing summaries by hand
Retrieval layer over your internal document base
New hires find the answer themselves instead of interrupting the one person who knows it
Phased migration from on-premise to hybrid, over two quarters
Capital expense becomes a monthly line you can forecast, defend, and cut
Services
Five things I get called for
Applied AI — and where it actually pays
Finding the places in your business where AI returns more than it costs, and saying so plainly when it doesn't. Pilots designed to answer a question, not to demo well. Agent factories included: how many agents, which systems they may touch, where a human signs off, and what one run costs.
Speech technology
Transcription and speech synthesis put to work: call analytics, voice interfaces, accessibility, media production. Model selection, tuning, and the pipeline around it.
Cloud, Kubernetes, Linux
Architecture and migration for infrastructure that simply has to keep running. An honest assessment of what you have, a plan for what you need, real numbers for both.
Development you can plan
Architecture, sizing and technical supervision for software other people build — mobile applications included. How hard the thing really is, what it will cost, which parts are risky, and whether the plan survives contact with a team. I don't take the keyboard.
The arithmetic
Run your own numbers
Idle capacity is the cheapest money on the table and the least often counted. Put in what you actually run — this is your arithmetic, not my estimate, and it stays in your browser.
Paid for nothing, per year
per month
The number is only the ceiling — some of that capacity is headroom you chose on purpose. Telling the difference is most of the work.
Stack
What the advice is built on
Seventeen years inside global technology companies — enterprise infrastructure, cloud platforms and machine learning, on the side that has to deliver what was sold. Long enough to know which promises hold after the contract is signed.
Infrastructure
- Public cloud platforms
- Kubernetes & containers
- Linux
- Networking
- Capacity & cost planning
Machine learning
- Speech-to-text
- Speech synthesis
- Speaker diarization
- Fine-tuning on domain data
- Evaluation & benchmarking
Applied AI
- Agent factories & orchestration
- Use-case discovery
- Retrieval over private data
- Agents & automation
- Build-versus-buy analysis
- Unit economics
Engineering
- Python
- Mobile applications — iOS, App Store
- APIs & integrations
- Data pipelines
- Proofs of concept
How it is built
What to keep, what to rent
The same stack, split by one question asked layer by layer: does owning this buy you anything? Hardware and the workloads on top usually stay. The middle usually does not.
On-premise
Cloud
-
kept Workloads the only layer your customer sees
-
rented Autoscaling capacity that answers to demand
-
rented Kubernetes control plane, then the nodes
-
rented Storage and its backups, which are not the same thing
-
kept Network routing, peering, firewalls
-
kept Linux the boring, load-bearing part
-
kept Bare metal racks, power, hands
Four layers kept because they have to be yours, three rented because owning them buys nothing.
About
Who you would be working with
Anton Gostev
Consultant, entrepreneur, engineer
My name is Anton Gostev, and I spent seventeen years inside large technology vendors: eleven at Microsoft, three at Yandex International, two at IBM and one at Dassault Systèmes.
The work always had the same shape — standing between an engineering team and a customer who has to sign, and making sure both were describing the same thing.
Today I do it on my own, through two registered practices: one in Serbia, one in Spain. You talk to the person who does the work, because there is no one else to hand it to.
Where those years went
- Microsoft — 11 years
- Yandex International — 3 years
- IBM — 2 years
- Dassault Systèmes — 1 year
How an engagement runs
Three steps, in this order
Step 01 — Assess
Find out what is really there
A week or two: what you run, what it costs, where it hurts, and what the team already knows but hasn't been asked.
Step 02 — Plan
Decide what is worth doing
Architecture and a costed roadmap that separates what to do now, what to defer, and what to drop entirely.
Step 03 — Execute
Stay until it works
Hands-on, alongside your team, or reviewing a vendor's work from your side of the table. Whichever the situation calls for.
Radar
What I'm reading this week
Updated automatically · 25 Aug 2026
Infrastructure, applied AI and the security news that actually changes what you should do on Monday. Headlines link straight to the source.
Spyre-Accelerated Retrieval-Augmented Generation on IBM LinuxONE: A Cloud-Native Architecture for Secure…
Running large language models inside enterprise environments has always bumped up against a practical wall: the data lives in one place, the AI horsepower sits somewhere else, and moving…
Read at source →
OpenAI and Anthropic in price war as Chinese AI rivals gain ground
US groups release cheaper models after new challenges to their trillion-dollar ambitions.
Read at source →
OpenAI Jalapeño: Better than Nvidia Blackwell
https://www.bloomberg.com/news/articles/2026-08-25/openai-cl... , https://archive.ph/yCTrr
Read at source →
OpenAI's upcoming Jalapeño chip looks like it'll be an inference beast
128 chips, 1.7 exaFLOPS, and 27 TB of HBM give Altman and crew a leg up over Blackwell, and maybe even Rubin
Read at source →
There Is No Neutral Harness: Modern LLM Leaderboards Are Manufactured by Config-Fragile Items
Multiple-choice benchmarks fix the questions and the correct answers, but not the harness: the order of the options, the wording of the prompt, and whether a language model's answer is read…
Read at source →
PBS station fears losing 50TB of data after being ghosted by cloud storage provider
"We don't have access to the data on the hardware/servers," Iron Mountain told Ars.
Read at source →
A curated digest, not a mirror: headline, source and a one-line summary, with the link going to the publisher.
Contact
Tell me what's not working
A first conversation costs nothing and usually takes half an hour. If I'm not the right person for it, I will say so and point you at someone who is.
What the first call looks like
- I don't sell anything on it — not on the first call, and not on any later one. I listen, and I ask questions.
- What I'm after is not only the problem, but where it came from: what caused it, or what it is a consequence of.
- And what you have already tried. The experience a team picks up while failing to fix something usually decides what will actually work.
- info@2cloud.me
- Phone
- +381 62 9326 061
- Belgrade
- Bulevar Milutina Milankovića 9ž, Belgrade 11070, Serbia
- Barcelona
- Pl. d'Europa 50-52, 08902 L'Hospitalet de Llobregat, Barcelona, Spain
- Working in
- English · Srpski · Español