Vladislav Ivanov

Sofia, Bulgaria · My AI lab

I test AI myself, then put it to work with your team.

Helping you choose the right AI tools for your processes.

I work with leadership teams one workflow at a time, with a person reviewing the result. More than a decade in ERP and blockchain came first.

See the labEmail me
New in my labLocal vs cloud · 10/10 checks passed by the guide my local model wrote

AI in Practice · my lab

3 experiments2 ongoing

New AI capabilities, tried on practical work.

Experiments, not client work: what worked, what didn't, what it took. This site is one of them, and it changes as I try new things.

EXP-003Local vs cloudOpen as a page
Replay · recorded final

My local model vs the best of build.nvidia.com: a Sofia guide in Bulgarian

Capability
A local open model against the best models NVIDIA's build.nvidia.com catalog offers on a free trial, on the same tasks in Bulgarian.
In practice
One brief: a printable one-page guide to a day in Sofia, in Bulgarian, with its own drawing, a day ruler, seven stops and prices. Ten cloud models went through four qualifying gates; four wrote the guide against my local DeepSeek V4.1 Flash.
Limits
One day of tests, one run per cloud guide, scored by a script; the notes are an AI reviewer's read. Only my local guide is shown: the trial terms do not allow sharing what NVIDIA's hosted models write.

Replay of the final · Sofia guide in Bulgarian · 8 Nov 2026

Cloud models tried
10on build.nvidia.com
Qualified
4passed all four gates
Local guide
10/10the only full score
Priced in euro
1 of 5Bulgaria’s currency since 2026
Time to write the whole guide, thinking included15:08 / 15:08
  1. 01DeepSeek V4.1 FlashLocal · DeepSeek5:1210/10€ euroNatural Bulgarian, prices in euro, correct metro changes. Some walking times are too short.
  2. 02Nemotron 3.5 Super VLbuild.nvidia.com · NVIDIA2:189/10лв levaFastest and tidy, but prices in leva and word slips, one of them vulgar.
  3. 03Nemotron 3 Ultrabuild.nvidia.com · NVIDIA7:139/10лв levaReal places, but text spills out of the cards and a metro line is wrong.
  4. 04Mistral Large 4build.nvidia.com · Mistral AI15:089/10лв levaThe best cathedral drawing, but most stops lack transport and prices.
  5. 05Step 5 Previewbuild.nvidia.com · StepFun2:448/10лв levaThe strongest type, but no day ruler and some invented places.

Localbuild.nvidia.comWaiting in a queueCheck passed

The local guide

The one-page Sofia guide my local model wrote: a line drawing of Sofia, a day ruler from 09:00 to 21:00 and seven stops, in Bulgarian, on near-black.
DeepSeek V4.1 Flash, local5:12 · 10/10
Download the PDF
Read the brief

A one-page print guide in Bulgarian, „Един ден в София“ (A day in Sofia), for someone in the city for the first time, getting around on foot and by metro on an autumn weekday.

The page is near-black #0A0A0B with a faint 24 px grid of #F2F2F0 at 4 % opacity. At the top, a mono caption in caps, letter-spaced: „СОФИЯ · 42.6977° N · 23.3219° E“. Below it, a line drawing you make yourself in inline SVG, the full width of the page and about a third of its height: the domes of the Alexander Nevsky Cathedral in front of the long ridge of Vitosha, in 1.5 px off-white strokes, with one signal-blue #8AB4FF dot where the sun sets behind the mountain. Then the title „Един ден в София“ in large type with tight tracking, off-white #F2F2F0, and one muted line in #A6A6AB under it that sums up the day.

The middle is the day as a ruler, like the shutter scale on a camera: a thin tick scale from 09:00 to 21:00, a tick every half hour in #34343A, the hours in mono under it, and seven blue markers where the stops fall. Under the ruler, seven minimal cards in two columns: surface #111113, 1 px #232326 border, 12 px radius. Each card shows the time large in blue tabular numerals, the place in off-white, one sentence on what to do there, and a mono line with how to get there (on foot, the metro line or the tram number) and an approximate price. Real places only, in an order that makes sense on foot.

At the bottom, „Три съвета за днес“: three short practical tips in muted text, split by hairlines, and a mono footer line: „СОФИЯ · ЕСЕН 2026“.

One complete HTML file, nothing before or after it. No JavaScript and nothing loaded from the internet: no web fonts, no images, no external links. Fonts: "Helvetica Neue", Helvetica, Arial for text and ui-monospace, Menlo for captions. @page { size: A4; margin: 0 }, and everything fits on one A4 page. No gradients, no shadows, no emoji, one accent colour. All visible text in Bulgarian.

Only my local guide is shown: NVIDIA’s trial terms do not allow sharing what its hosted models write.

An afternoon on build.nvidia.com

  1. 01Good

    A free key in minutes. One standard chat API with 84 models behind it, so trying another model is a one-line change.

  2. 02Snag

    NVIDIA’s hosted DeepSeek V4.1 Flash, the model I run locally, sent nothing for five minutes, then HTTP 504. Its forum had the same reports for days.

  3. 03Mixed

    Nemotron 3.5 Lightning answered in under a second; Llama 3.1 8B answered HTTP 410, retired on 26 Aug 2026.

  4. 04Fine print

    Trial use only, not production. Prompts may be kept 30 days and used to improve NVIDIA’s models, and what the models write may not be shared.

  5. 05Snag

    Kimi K3.1 Pro, released 4 Nov 2026, is playground only: “API access for this model is not available yet.” Kimi K2.6 is listed but locked for my account.

  6. 06Snag

    Qualifier, 10 models in parallel: 3 sent nothing within four minutes; Mistral Large 4 waited 2 to 3.5 minutes in a queue before every answer.

  7. 07Good

    Nemotron 3.5 Super VL wrote at about 145 tokens per second and reads images. Nothing here was quicker.

  8. 08Result

    4 qualifiers wrote the guide in 2:18 to 15:08. None priced the day in euro.

The qualifier · 10 models, four gates: available, correct, Bulgarian, stamina

  • Nemotron 3 UltraFinal
  • Mistral Large 4Final
  • Nemotron 3.5 Super VLFinal
  • Step 5 PreviewFinal
  • MiniMax M3.1 FlashMissed Sofia facts
  • Nemotron 3.5 LightningMissed Sofia facts
  • DeepSeek V4.1 FlashNo answer
  • Gemma 4 31BNo answer
  • GLM 5.3 FlashNo answer
  • Kimi K2.6Locked

A replay of the recorded final on 8 Nov 2026: the clock runs fast, but every bar keeps its real time. Ten checks were scored by a script, the same way for all; the notes are an AI reviewer’s read, not a native speaker’s score. Nothing here runs live.

Goal

Is my local model good enough for real work in Bulgarian, and what can NVIDIA's free model catalog add?

How it ran

  1. One brief for every model: a printable one-page guide to a day in Sofia, in Bulgarian, in this site's design. It never names a currency.
  2. Ten models from build.nvidia.com ran quick tasks first: a short answer, JSON from an email, one fact in a long log, a 700-word guide, a photo and five Sofia facts in Bulgarian.
  3. Four passed all four gates (available, correct, Bulgarian, stamina) and wrote the guide against my local DeepSeek V4.1 Flash.
  4. A script scored every page on ten checks, the same way for all, and printed it to an A4 PDF.

Misses and limits

  • NVIDIA's hosted DeepSeek. The model I run locally never answered there: nothing for five minutes, then HTTP 504.
  • Locked and retired. Kimi K3.1 Pro is playground only, Kimi K2.6 is locked for my account and Llama 3.1 8B is retired.
  • My own check. The real-places list missed landmarks one model used; I widened it and scored every guide again.
  • What a script cannot see. Overlapping text, wrong words and wrong metro lines need a reader; the notes are an AI reviewer's read, not a native speaker's.

Nice to have

  • How my local model runs, as an experiment of its own.
  • The final again, if Kimi K3.1 Pro opens its API.

Want to know which models are worth running for your team, and which ones only look good in a demo? Let’s talk.

Email meLimeShift.ai (opens in a new tab)LinkedIn (opens in a new tab)

New experiments are added here as they run. Have a question you want tried? Email me one line.

How I help

Three ways I work with your team.

I start with one process your team already runs and choose the AI tools that fit it. Then we work on it together until it runs every week, with a person reviewing the result. I do this through LimeShift.ai, which grew out of AI work inside LimeChain.

ERP and blockchain taught me to look at the process first and the tool second. I still test the models and agents myself. I finished an Ironman the way I put AI to work: one steady step at a time.

  1. 01

    One workflow at a time

    We start with the AI tools your team already knows, then take one process that repeats: reporting, sales ops, support or engineering. It gets approved data, an owner, a metric and a person who reviews the result. When it works, we take the next one.

  2. 02

    Choosing the right tools

    Which model, which agent setup, build or buy, and whether your data is ready. I test these myself, so you get a straight answer, including the risks nobody put in the slide deck.

  3. 03

    Staying for the long middle

    The first version is the easy part. I have built and run delivery teams since 2017, so I stay for what comes after: owners, hiring, delivery and the long middle of a build.

The method, every time

  1. One workflow
  2. Approved data
  3. AI processing
  4. Human reviewGate
  5. Measure
  6. Expand

Experience · Career arc

  1. 01about 2013 to 2017

    ERP

    Seeing how a business translates into a working system, and where the money actually moves.

    • Led ERP implementations across manufacturing and logistics
    • Rebuilt order-to-cash processes for mid-sized enterprises
    • Microsoft Dynamics developer at XAPT CEE: Dynamics AX and Dynamics NAV (Navision), including Bulgarian localization
  2. 022017 to now

    Blockchain

    Learning when new technology earns a place in a real process, and when it doesn't.

    • Co-founded LimeChain in 2017 with Nick Todorov, Chris Veselinov and George Spasov; Head of Development since founding
    • LimeChain: 50+ people, 100+ projects, zero external funding, Sofia HQ and a London office
    • Engineering lead on blockchain and smart-contract projects for protocols and enterprises
  3. 032026 to now

    AI · LimeShift

    Choosing the right tools from everything I have already seen, and making them part of how a team actually works.

    • LimeShift.ai is a governed AI workflow practice that grew out of AI work inside LimeChain
    • Proven inside LimeChain across 5 business functions
    • Signature line: Most companies do not need more AI experiments. I run those in my lab. They need AI to become part of how their team works.

Most companies do not need more AI experiments. I run those in my lab. They need AI to become part of how their team works.

Two rules I work by.

  • On adoption

    Use AI on your own work before you roll it out.

  • On trust

    I show up with the work, keep my promises and put the client’s outcome ahead of my own.