Machine Learning Engineering
Custom models built around your reality, not a vendor’s demo dataset. We design, train and distil models that survive contact with messy production data — and we stay until they earn their keep on your P&L.
/01AccessBG — Artificial Intelligence Studio · Sofia, Europe
Intelligence, made accessible.
We are an artificial-intelligence studio born in Sofia and built for Europe. We take the most consequential technology of our lifetime and make it usable, dependable and quietly beautiful — turning the data you already own into systems that see, read, predict and decide.
The studio
Most companies don’t have an AI problem. They have an access problem. The intelligence is already in their data — locked behind formats, silos and doubt. Our job is to hand them the key.
AccessBG began in 2014 as three researchers and one stubborn conviction: that artificial intelligence should not be a privilege of trillion-dollar platforms. It should belong to the energy company in Plovdiv, the hospital network in Varna, the logistics fleet crossing the Balkans at 3 a.m. — to anyone with real problems and real data.
Today we are a senior team of machine-learning engineers, research scientists, data craftspeople and product designers operating from Sofia Tech Park. We have shipped 87 models into production across energy, healthcare, law, manufacturing, retail and logistics — systems that quietly make several billion predictions a month while their owners sleep.
We call our practice applied intelligence: research-grade rigor, shipped with the discipline of industrial software. No demos that die in slides. No black boxes you’re forbidden to open. Just intelligence, made accessible — достъпен, as we say at home.
0+
Years of applied research
0
Models in production
0+
Teams empowered
0B
Predictions / month
Capabilities
Six disciplines, one craft. Each capability below is a full team — researchers who read the papers, engineers who ship the systems, and the unglamorous middle where the two actually meet.
Language Models & Agents
Retrieval-augmented systems, domain fine-tuning and agentic workflows that read your contracts, answer your customers and draft your reports — fluently, in Bulgarian, English and twenty-two other languages, with citations you can verify.
/02Computer Vision
Eyes that never blink. Defect detection on production lines, document OCR at archive scale, safety monitoring, 3D scene understanding — vision systems trained on your world and tuned for the edge cases that actually cost you money.
/03Predictive Analytics
Tomorrow, quantified. Demand forecasting, churn and risk scoring, dynamic pricing — probabilistic models with honest uncertainty bands, so your planners argue about strategy instead of arguing with the spreadsheet.
/04MLOps & Infrastructure
A model is a promise; infrastructure keeps it. Versioned pipelines, drift monitoring, automated retraining and rollback — on your cloud, on-premise, or on a device the size of a matchbox bolted to a factory wall.
/05Strategy & Governance
AI you can defend in a boardroom and a courtroom. Opportunity audits, build-vs-buy roadmaps, EU AI Act readiness and executive training — so your organisation adopts intelligence deliberately instead of accidentally.
/06Why AccessBG
Anyone can rent a model. Few can make one trustworthy. These are the six commitments we refuse to negotiate — the reason teams stay with us for years, not quarters.
①
Research-grade, production-first
We read the papers on Sunday and ship the pipelines on Monday. Every idea earns its place with a benchmark, then earns its keep in production.
②
Models you can audit
Explainability is not a feature request — it’s the default. You get the training data lineage, the evaluation suite and the failure modes, in writing.
③
European by design
GDPR-native architectures and EU AI Act conformity from the first sketch. Your compliance officer will be suspicious of how easy this was.
④
Senior people only
The team you meet in the pitch is the team in your repository. No bait-and-switch, no army of juniors learning on your invoice.
⑤
Your data stays yours
We train on your infrastructure when needed, sign what your lawyers draft, and walk away owning nothing but the lessons.
⑥
Outcomes over deliverables
We price against the decision we improve — forecast error, hours saved, defects caught — not against the kilograms of slides we produce.
The shape of the studio
Research depth 96%
Engineering rigor 94%
Data craft 91%
Product thinking 88%
Self-reported, ruthlessly argued over at the studio retreat. The missing percentages keep us humble.
Selected works
Six systems, six industries, one habit: measurable before beautiful. Names are real, clients are under NDA — the numbers were audited by people paid to doubt us.
Helios
A national grid operator was burning money on imbalance penalties. Helios fuses weather physics with gradient-boosted ensembles to forecast load and solar feed-in at 15-minute resolution — cutting day-ahead error by 38% and paying for itself in eleven weeks.
Lexis
A pan-European law firm drowning in data rooms. Lexis is a retrieval-augmented language engine that reads 120,000 pages an hour, flags non-standard clauses and answers in lawyer — every claim pinned to a paragraph-level citation. Diligence that took weeks now takes an afternoon.
Iris
On a packaging line moving four units a second, the human eye taps out by lunch. Iris watches every unit, every shift, catching surface defects at 99.4% recall on cameras that cost less than one month of scrap. The line manager calls it “the colleague who never blinks.”
Pulse
Emergency departments don’t fail from lack of care — they fail from lack of foresight. Pulse forecasts patient inflow 72 hours ahead from weather, events and history, letting a five-hospital network roster staff where the wave will land. Average waiting time fell 27% in one winter.
Vox
Bulgarian is spoken by nine million people and understood by almost no voice assistant. Vox fixed that: a speech-and-language stack fine-tuned for Bulgarian and 23 other European languages, now answering 64% of a telecom’s support calls end-to-end — politely, even at 2 a.m.
Atlas
A 400-truck fleet, ten thousand constraints, one impossible spreadsheet. Atlas re-plans routes every twenty minutes against live traffic, dock slots and driver-hours law — shaving 19% of kilometres and an entire dispatch night-shift of stress. The planet noticed too: 1,900 fewer tonnes of CO₂ a year.
The method
Six movements, rehearsed across a decade of engagements. The order never changes; the tempo adapts to you. Most clients see a working prototype before the first invoice clears.
-
01
Discover
One immersion week inside your operation. We interview the sceptics first, map where decisions are made, and weigh the data you actually have against the data everyone assumed you had.
-
02
Frame
We translate ambition into a falsifiable hypothesis: one decision, one metric, one baseline to beat. If we can’t define what “better” means in numbers, we don’t proceed — and we tell you so for free.
-
03
Prototype
Within three to five weeks you hold a working system fed by your real data — deliberately rough, deliberately honest. This is where conviction is built or the project is mercifully ended.
-
04
Train & evaluate
The long middle. Architectures are tuned, data is cleaned until it confesses, and every model faces an adversarial evaluation suite — bias probes, drift scenarios, red-team prompts — before it earns a version number.
-
05
Integrate
Intelligence is useless in a notebook. We wire the model into your ERP, your line, your call centre — with fallbacks, audit logs and a UI your least patient colleague will actually use.
-
06
Operate & evolve
Models age like bread, not wine. We monitor drift, retrain on schedule, and hand your team the keys at the pace they want them — until the system is theirs and we’re just the people they call with good news.
Voices
The people who signed the invoices, in their own words.
We had been promised AI three times before — twice by global consultancies with logos bigger than their results. AccessBG was the first to ask what decision we were trying to improve. Eight months later our imbalance costs are down by a third and my board asks about the model by name.
They flew to Milan, sat with our dispatchers for a week, and only then opened a laptop. The routing engine they built does in twenty minutes what our night shift did in six hours — and the team trusts it, which is the part nobody else managed.
In healthcare you cannot deploy what you cannot explain. AccessBG delivered the prediction model together with its full evaluation dossier — our clinical board approved it in a single sitting. Waiting times dropped within the first quarter.
Honest to the point of discomfort. They killed one of our pet projects in week two because the data couldn’t support it — then redirected the budget to a forecasting system that now steers €40M of seasonal inventory. That honesty is why they’re still here, four years on.
Let’s give you access
Tell us about the decision you wish you could make better. Three sentences are enough — we’ll reply within one working day with questions, not a sales deck.