Education and EdTech

AI that keeps children in school and teachers teaching

Q: We are a 180-student school. Can we afford any of this?

Probably not directly, and we would rather say so on the first call. A school that size pays on the order of €310–385 a year (600–750 лв. before the changeover) for its e-diary, partly reimbursed per student through National Programme „ИКТ в системата на предучилищното и училищното образование“, and no bespoke engagement fits inside that. Our buyers are municipalities, РУО, МОН programme teams, publishers and e-diary vendors — organisations serving hundreds or thousands of schools at once. If you are one school with one painful process, ask for the shortest possible answer and we will give it without an invoice.

Q: We run a public procurement. What have you actually delivered in Bulgarian education?

The digital dictionary at beron.mon.bg — contracting authority the Ministry of Education and Science, a public deployment you can open and inspect yourself without asking us for anything. The reference-grade specifics a ЗОП evaluation needs — year, contract or programme reference, scope, and a удостоверение за добро изпълнение if the authority issues one — we supply in writing on request and enter in the ЕЕДОП; we do not print figures on this page that we cannot evidence in a procurement file. Outside education: a European car marketplace handling 300,000+ listings, an NLP module that answers trading questions across email and Instagram with a person approving every reply before it sends — about 85% less manual typing on repeat questions, measured by the client — and email-marketing automation that cut a beauty brand’s outreach time by 60%, also client-measured. That is the complete list. No education awards, no 1EdTech conformance certification yet, no roster of school logos — if the bid requires those, we are the wrong bidder and you should score us accordingly.

Q: We already run Школо. Do we have to replace it?

No, and you should not. Bulgaria has roughly eight competing e-diaries — Школо, АдминПлюс, Сиела, Е-дневник, Феникс and others — plus the НЕИСПУО Дневници module, and the diary is where teachers already live every morning. We read its exports, launch from it over LTI 1.3, and write files back in the formats it and НЕИСПУО expect. Anything that asks a 55-year-old teacher to learn a second daily tool has already failed, whatever the feature list says.

Q: Will our students’ data be used to train a model?

No. Retrieval is not training: approved textbooks and учебни програми are indexed, the model is not fine-tuned on student work, and identified data stays on infrastructure you control. Where fine-tuning genuinely helps, it runs on de-identified or synthetic material with the derivation documented in the DPIA. The ЛОД is retained 50 years under Наредба № 8 as a statutory obligation under GDPR Art. 6(1)(c) — that is a retention duty, not a licence to reuse the record for anything else.

Q: The research says AI makes students learn less. Why would we buy this?

Some of it does, and the strongest study deserves to be taken seriously: a US panel of 3.2 million ALEKS interactions found that study time on AI-susceptible maths problems fell 26.9% cumulatively for college students and 31.3% for high-schoolers; that the time reduction did not appear in proctored practice, where AI was unavailable; and, separately, that performance on proctored retention items — the check that AI use had not been learning — declined 25% in the odds of a correct response. Not a Bulgarian cohort, and we hand over the full reference in writing. That is an argument against unsupervised answer generation for students, not against every application. It is also why our default work targets teacher workload and administrative burden, where the worst failure mode is a wasted draft rather than an unlearned skill.

Q: What do we own at handover, and can we leave?

Assessment content in QTI 3.0, packaged courses in Common Cartridge, rosters and enrolments in OneRoster 1.2, lexical data in TEI Lex-0 with an OntoLex-Lemon projection, learning events as xAPI into an LRS you own, and the pipelines, curriculum URIs and evaluation harness in your repository. The export endpoints are named in the statement of work and demonstrated during acceptance testing, not at termination — and for work for a state authority, the custom source code in the state repository at github.com/governmentbg, in the form ДАЕУ acceptance requires. Handover includes training a named person on your side to re-run and re-score the whole thing without us.

We work across the whole span of a school system: from the one certificate that expires on the one machine the director signs from on the 3rd of the month, and the ЛОН that appears twice because a transfer was typed with a different ЕГН, to the 30,000-resource national repository a teacher cannot search for next Tuesday’s lesson. Most of it is unglamorous plumbing — validation rules out of Наредба № 8, duplicate reconciliation, curriculum URIs, alt text — with a model at the end of it, if one is needed at all. The same calendar runs in the 1,829 independent kindergartens, where the return is Списък-Образец № 2 and the absence trigger is more than three days rather than five school hours.

Book a call See our projects →

30 minutes with the engineer who would do the work — not a salesperson. No obligation, and you keep whatever we work out on the call.

Teacher working through a lesson on a laptop with two students in a classroom

The numbers this sector is run on

60–80%: Schools whose Списък-Образец № 1 clears validation first timeOur estimate, from conversations with school administrators and РУО staff — no national first-pass figure is published, which is itself part of the problem. If you hold the real number for your region, it replaces ours on the first call. On the same evidence a first submission typically throws 5–40 errors, dominated by staffing arithmetic against the задължителна норма преподавателска заетост and curriculum-plan mismatches. Above 95% is achievable, and it is visible in the РУО’s September escalation volume.
0.5–5%: Absence records needing manual correction on importOur estimate from the imports we have seen, not a published statistic. Above 5 unexcused school hours in a month, the monthly помощ for that child is withheld under the Закон за семейни помощи за деца — by a decision of the директор на дирекция „Социално подпомагане“, on data the school confirmed, appealable under АПК. Each import error is therefore an administrative act against a household, and the evidence you will be asked for is per-absence: who recorded it, who excused it, and when.
60–85% vs 15–40%: Parents monthly active in the diary: urban versus ruralThe most honest adoption number for anything parent-facing, and it comes from directors rather than from a national survey. They report 60–70% of parents in some village schools never logging in at all because they work abroad, and, in schools with a high concentration of students from vulnerable groups, 10–15% who can actually use the diary. A national average here means nothing.
5–20%: Teachers weekly active on a national content platformWithout a mandate and without curriculum integration. Our estimate — no adoption report for Дигитална раница has ever been published, which is the first thing we would fix. When resources are embedded in the тематично разпределение the band moves to 35–60% — the difference is not the product, it is whether using it is part of the required work.

What we build

School administrator reviewing an academic calendar beside a laptop

The compliance calendar, treated as an engineering problem

Списък-Образец № 1 validated in two passes — carry-forward structure from 1 September, live data from the day enrolment and the паралелка заповед are fixed — against a rule engine encoding Наредба № 8 and the framework curricula: assigned hours against the задължителна норма, group splits against enrolment, every student in exactly one паралелка with a valid ЛОН. Absence campaigns orchestrated across the 1st to the 4th, including certificate expiry and signing-agent health on the specific machine the director signs from.

Errors returned as a ranked worklist naming the sub-tab and the field, with a diff against last year’s approved return
Signatures modelled for the director, the chief accountant and the first-level spending officer — a workflow with one approver is already wrong
Shadow mode for a full cycle before anything we build sits in the critical path
A РУО-side view of every school’s validation state before 20 September, so the approval window on 21–25 September is triage rather than discovery

Technician checking cabling on an on-premise server rack

Bulgarian models that run on hardware you control

BgGPT ships openly in Gemma-2 and Gemma-3 variants with GGUF and GPTQ-W4A16 quantisations, so Bulgarian-competent inference runs in-country on your own GPUs. We separate the model that touches identified student data, which stays on-premise, from offline content generation, which may use an external API precisely because no personal data reaches it. The evaluation set comes from real НВО and ДЗИ items.

Retrieval over МОН-approved textbooks and учебни програми, with the source unit cited in every answer
Results published per subject and grade band, including the ones that fall below the bar
DPIA, ROPA entry, sub-processor register and data-flow diagram before code, not after procurement
The Art. 27 fundamental rights impact assessment that falls on you as a public-body deployer of an Annex III system, drafted with you rather than left to your legal directorate

Teacher selecting teaching material on a laptop in a staff room

Content a teacher can actually find on a Monday

A CASE-style machine-readable model of the Bulgarian учебни програми gives every curriculum outcome a stable URI. Resources are ingested with OCR where needed, classified against that taxonomy by a model constrained to it, tagged with LRMI and schema.org metadata, and clustered for near-duplicates by embedding. Low-confidence classifications go to a subject expert. It launches over LTI 1.3 from the diary the teacher already opens.

Curriculum coverage becomes a number: units by subject and grade with at least one approved, tagged resource
OneRoster 1.2 for rosters, QTI 3.0 for items, Common Cartridge 1.3 for packaged courses
No new login and no new tab — 31.4% of Bulgarian teachers are 55 or older, and that is a design constraint

Student using large accessible text on a laptop with headphones

Accessibility generated at ingest, not retrofitted

EN 301 549 applies from the first asset, so alt text, MathML and captions are produced inside the pipeline rather than bought back later across tens of thousands of files. Bulgarian alt text is generated from the image and its surrounding lesson, then reviewed wherever it carries instructional load. Equation bitmaps become MathML with a mathematician in the loop, and captions are gated on word error rate.

axe-core and Lighthouse surface only 30–40% of real defects; the rest needs keyboard and screen-reader testing
Conformance tracked per asset, so the accessibility statement is generated from evidence
WCAG 2.2 criteria that bite in school UIs: 2.5.8 target size, 3.3.7 redundant entry, 3.3.8 accessible authentication

Worked examples

These are designs, and the ranges we would contract against, drawn from published sector data — not Palamed results. The four systems we have actually delivered are on case studies.

Списък-Образец № 1 pre-flight validation

Problem

The return is assembled by hand across six sections — Институция, Групи и класове, Персонал, Деца и ученици, Учебен план, Номенклатури — and the errors surface only when НЕИСПУО rejects it, days before 20 September. Staffing arithmetic is where most schools break: teaching hours divided by the задължителна норма to derive FTEs. Neither the director nor the chief accountant can КЕП-sign a return that does not reconcile. And the school’s deadline is not the end of it — the РУО must approve each return by 25 September, so every late resubmission lands inside a five-day window on a directorate reviewing hundreds of schools at once. The same cycle repeats for the mandatory updates approved by 5 January and at the start of the second term.

Approach

A rule engine over the e-diary export and last year’s approved carry-forward, encoding Наредба № 8 and the рамкови and типови учебни планове: weekly hours × weeks against declared totals, group splits against enrolment counts, every паралелка mapped to a plan, every teacher’s assigned load against the contracted norm, every student in exactly one class with a valid ЛОН. Output is a ranked worklist naming the exact sub-tab and field, plus a diff against the approved prior-year return so unexplained deltas surface on their own. It runs in two passes. From 1 September it validates everything that carries forward — curriculum plans, staff loads, номенклатури — so structural errors are cleared before the data that cannot exist yet arrives. From the day enrolment and the паралелка заповед are fixed, usually 15 September, it revalidates continuously against live diary data, so the last five days are spent on genuine disagreements rather than on arithmetic you could have caught two weeks earlier.

What we would target

First-pass acceptance moves from roughly two-thirds of schools toward the 95% band, and the РУО’s September escalation volume falls with it. Administrator effort on the return drops from several days to under a day, and what is left are real disagreements about the plan rather than arithmetic.

The absence campaign, from diary to КЕП to АСП

Problem

Between the 1st and the 4th of every month the school’s НЕИСПУО administrator imports absences, reconciles them, and gets the director to sign with a КЕП through a signing application installed on one specific machine. Drivers break after a reinstall, certificates expire unnoticed, class teachers backdate excusals after the file has been cut — and above five unexcused hours in a month the monthly помощ for that child is withheld under the Закон за семейни помощи за деца, by a decision of the директор на дирекция „Социално подпомагане“ on data the school confirmed, appealable under АПК.

Approach

An orchestration layer that pulls absences from whichever diary the school runs, applies excusal rules and flags medical notes arriving after the file was generated, diffs the school’s record against the АСП-loaded campaign so only disputed students need review, and monitors signing-agent health and certificate expiry on the 20th rather than on the 3rd. Every change carries an immutable trail of who edited which absence and when. НЕИСПУО never deletes submitted data — corrections stay visible next to the original — so the audit design is a given, not a feature request.

What we would target

Signed submission by day 4 across the whole estate instead of a chased rural tail, disputed АСП records reviewed in minutes rather than hours, and a defensible evidence trail when a parent appeals the withholding under АПК — what you will be asked for is per-absence: who recorded it, who excused it, and when. The import error rate becomes a tracked number instead of a monthly discovery.

Dropout risk wired into the Механизъм за обхват

Problem

The coverage mechanism dispatches joint municipal, school and police teams to addresses, but it fires on children already registered as out of school. Скрито отпадане — enrolled on paper, absent in reality — is invisible, because per-student delegated funding rewards keeping the headcount intact and under-recording absences. Early leaving from education and training — the Eurostat 18–24 measure — runs at 17.7% in rural areas against 3.4% in cities, against a national 8.2%. That is the outcome we are trying to move. What the model actually predicts is different and narrower: within-year disengagement among currently enrolled students in НЕИСПУО, which runs at a few percent nationally. We report both, and we never evaluate the classifier against the ESL rate.

Approach

A model on absence trajectories, grade trends, transfer history, ЦДО participation and support-need flags already held in НЕИСПУО, calibrated with isotonic regression and thresholded to a recall floor rather than the default 0.5 — the published system tuned to 0.387. Subgroup true-positive rates are audited by settlement type before deployment, because aggregate accuracy hides rural underperformance. Alerts per case worker are capped at 15–25 a week or the list stops being read. Explanations come at three levels: global feature importance for policy, marginal effects for the РУО, per-student counterfactuals for the class teacher.

What we would target

The published system reached precision 0.751 at a 27% positive rate. Bulgarian within-year leaving among enrolled students runs closer to 3%, and precision falls with prevalence: at a recall floor of 0.85 you should expect roughly one true case in five to eight flags, not one in four. We size the alert budget to that, which is why the cap is 15–25 per case worker rather than a ranked national list. We will state the expected precision on your data before you sign, from a backtest on three prior НЕИСПУО cohorts. The by-product is often worth more than the model: cross-checking flagged students against real attendance makes systematically under-recorded absences visible to the РУО for the first time.

A Bulgarian teaching assistant that cites the textbook

Problem

General chatbots answer curriculum questions in Bulgarian with confident errors, cite nothing, and cannot be pointed at the specific textbook a school adopted under Наредба № 10. Teachers will not stake a lesson on output they have to verify line by line, and anything that evaluates learning outcomes or steers the learning process sits in AI Act Annex III point 3(b). The hard part, though, is not retrieval. That approval confers no reuse rights — the textbooks are the publishers’ copyright, and indexing them for a national assistant is a licence negotiation with Просвета, Клет, Изкуства, Педагог 6 and Архимед before it is an engineering task. We scope the corpus in three tiers: учебни програми and МОН-published material, which are free to index; repository resources whose ingest terms already permit it; and publisher textbooks, which need a written licence we help you draft and which we will not index without one.

Approach

Retrieval over an indexed corpus of approved textbooks, учебни програми and vetted repository resources, with every answer required to cite the source unit. Served by BgGPT — Gemma-3 12B or 27B in GPTQ-W4A16 — hosted in-country, so no identified student data crosses to a third-country API; a frontier model is used only for offline content generation where no personal data is involved. Scope is constrained explicitly: it drafts, explains and suggests, and it does not award a mark. Every generation is logged with its retrieved context, and the evaluation set is built from real НВО and ДЗИ items rather than translated benchmarks.

What we would target

Citation coverage above 95% and unsupported-claim rate under 2% on the curriculum eval set, p95 latency under 2.5 seconds, and a component we argue outside Annex III under Art. 6(3) — no automated output evaluates a learning outcome and the assistant does not profile students — with that assessment documented and registered rather than assumed. Where a subject fails the bar, we ship the ones that pass and name the one that did not.

Accessibility remediation across a national repository

Problem

A catalogue assembled from publisher PDFs, teacher slide decks and 3D scenes fails EN 301 549 nearly everywhere: images without alt text, mathematics shipped as bitmap images, video without captions, tables without headers, contrast failures in teacher-authored slides. Directive (EU) 2016/2102 requires a published accessibility statement, a working feedback mechanism and periodic monitoring — and hand-remediating tens of thousands of assets is not affordable.

Approach

An ingest pass that generates Bulgarian alt text from the image plus its surrounding lesson context and routes anything instructional or low-confidence to human review; converts equation images to MathML through OCR with a mathematician checking the output; produces Bulgarian captions with Whisper-class ASR behind a word-error-rate gate, because above roughly 10% WER captions stop aiding comprehension; and repairs document structure — heading levels, table headers, reading order, language tags. axe-core runs on everything, a manual keyboard and screen-reader audit runs on a sample, and conformance is recorded per asset.

What we would target

WCAG 2.1 AA across the catalogue at a fraction of the manual cost, with EN 301 549 V4.1.1 and WCAG 2.2 already scoped rather than deferred. The accessibility statement is generated from a per-asset record instead of drafted by a lawyer, and students with СОП — including those supported through the РЦПППО and the school’s own resource team — use the same resources as their classmates, which was the point of the exercise.

Systems we work with

We integrate with what you already run. If a platform below is missing, tell us — the pattern usually transfers.

Statutory systems, identity and signatures

НЕИСПУО (neispuo.mon.bg)Институции, Деца и ученици, Дневници, Потребители, Поддръжка
edu.mon.bgthe МОН identity domain and institutional SSO accounts
AdminPro and AdminLthe МОН desktop applications e-diaries interoperate with
ИСРМ (back2school.mon.bg)the Механизъм за обхват information system
ИСОДЗ (kg.sofia.bg)Sofia kindergarten and first-grade admissions
Националното електронно класиране за прием след VII клас (priem.mon.bg)the highest-load day in the education calendar and an Annex III 3(a) system, which is exactly why we keep the ranking deterministic and published
ЦКОКУОНВО and ДЗИ delivery and result publication
ЦАИС ЕОПwhere the bid is actually submitted
КЕП signing agents under eIDASB-Trust (Борика), StampIT (Информационно обслужване), Евротръст, ИнфоНотари — plus еВход, RegiX and Keycloak. Renewal windows and driver packages differ by provider, which is why certificate monitoring is per-QTSP rather than generic.

E-diaries, timetabling and school operations

Школо (Shkolo.bg)the largest private diary, auto-generates the НЕИСПУО absence return
АдминПлюс
Сиела e-дневник
Е-дневник
Феникс
The НЕИСПУО Дневници module itself
aSc TimeTables
Untis
FET
SELFIE and SELFIE for Teachers
DigCompEdu

Content, LMS and interoperability standards

Дигитална раница on edu.mon.bg
mozaBook
mozaWeb and mozaMap
e-Просвета and e-uchebnik.bg
Клет България
Изкуства
Педагог 6
Архимед
Moodle, Google Classroom, Microsoft Teams for Education with School Data Sync, Canvas, Open edX
LTI 1.3 and LTI Advantage
OneRoster 1.2
QTI 3.0
Common Cartridge 1.3
CASE 1.0
Caliper 1.2
H5P, SCORM 1.2 and 2004, cmi5, xAPI into Learning Locker or Veracity

Bulgarian language technology and accessibility tooling

BgGPT (INSAIT)Gemma-2 2.6B/9B/27B and Gemma-3 4B/12B/27B, GGUF and GPTQ-W4A16
Български национален корпус
BulNet
BulSemCor
BulTreeBank
CLaDA-BG
TEI Lex-0
OntoLex-Lemon
LMF (ISO 24613)
Lexonomy
Sketch Engine
Whisper and faster-whisper for Bulgarian ASR and captioning
axe-core
Pa11y
NVDA
JAWS
VoiceOver
MathML Core
EPUB Accessibility 1.1

What we design against

0.85: Dropout recall floor four to five months out, at one true case in five to eight flagsA published early-warning system on 451,852 students reached recall 0.867 at precision 0.751 and AUC-PR 0.888 on a cohort with a 27% positive rate, at a tuned threshold of 0.387 rather than 0.5. Equal-opportunity post-processing raised recall to 0.910 but dropped precision to 0.702 and moved the false-positive rate from 0.106 to 0.142. Precision scales with prevalence, and Bulgarian within-year leaving among enrolled students runs closer to 3%: at a recall floor of 0.85 expect roughly one true case in five to eight flags, not one in four. We size the alert budget to that, and we state the expected precision on your data before you sign, from a backtest on three prior НЕИСПУО cohorts. Not a Palamed result, not a Bulgarian cohort — we hand over the full reference in writing.
>95% / <2%: Citation coverage and unsupported-claim rate on curriculum answersThe bar we set for a Bulgarian assistant grounded in approved material. Measured on a held-out set built from real НВО and ДЗИ items and МОН-approved textbooks, never from translated English benchmarks. Reported per subject and grade band, including the bands that fail it. It is a target we design against and report against, not a result we are claiming.
20–60 min: Administrative time returned per teacher per weekOur estimate for document generation pre-filled from НЕИСПУО and diary data — no published Bulgarian figure exists. Anything claiming several hours a week is measuring different work. The baseline comes from a two-week time diary run before we build, not from a brochure, and the number reported afterwards is yours rather than ours.

Regulation and standards in scope

ЗПУО чл. 250 and Наредба № 8 от 11.08.2016 г. за информацията и документите (ДВ бр. 66/2016, amended repeatedly — at the time of writing most recently ДВ бр. 65 of 8.08.2025, in force from 1 October 2025; we re-diff the ordinance against our rule set after every ДВ publication and the changelog ships with the product) — the standard defining every module, register, document type and retention period. The ЛОД is kept 50 years in НЕИСПУО and also printed with numbered pages, signed and stamped, which is a data model with a paper obligation attached.
Regulation (EU) № 910/2014 (eIDAS) and the Закон за електронния документ и електронните удостоверителни услуги — every НЕИСПУО submission is confirmed with a qualified electronic signature by the director and the chief accountant, from a machine running a locally installed signing application. That local component is the single most common monthly failure point, and any design that ignores it will be discovered on the 3rd.
GDPR Art. 6(1)(c) as the basis for statutory school records, Art. 8 on children’s consent, Art. 30 ROPA, Art. 35 DPIA (mandatory for large-scale processing of children’s data), Art. 37(1)(a) mandatory DPO — read together with ЗЗЛД чл. 25в, which sets Bulgaria’s digital-consent age at 14, mirrored in the rule that e-profile credentials go to students from 14 and to legal representatives below that.
Regulation (EU) 2024/1689 (AI Act) Annex III point 3 — access and admission (a), evaluating learning outcomes including steering the learning process (b), assessing appropriate education level (c), monitoring prohibited behaviour during tests (d). Art. 4 AI literacy obligations have applied since 2 February 2025; the Digital Omnibus, in force from 27 July 2026, deferred most Annex III obligations to 2 December 2027. Art. 14 human oversight is the article that shapes the architecture. Art. 6(3) is the only exemption route, and its final subparagraph keeps any system that profiles natural persons high-risk whatever oversight surrounds it; Art. 26 sets deployer obligations and Art. 27 requires a fundamental rights impact assessment from public-body deployers.
Directive (EU) 2016/2102 transposed through the Закон за електронното управление, EN 301 549 V3.2.1 (WCAG 2.1 AA today, V4.1.1 expected to bring WCAG 2.2 AA), and Directive (EU) 2019/882 applicable since 28 June 2025 for commercial textbook and e-book platforms — alongside NIS2 as transposed through the Закон за киберсигурност, which reaches ministries and agencies as public administration entities rather than schools as education entities, and which makes your suppliers your problem: incident-reporting timelines, supply-chain security requirements and the security clauses we expect to sign as your processor.
Закон за електронното управление and the Наредба за общите изисквания към информационните системи (ПМС № 493/2016) — a project for a state authority is registered in the ДАЕУ portfolio and approved before procurement; custom source code is deposited in the state repository at github.com/governmentbg under the applicable ЗЕУ obligations; the system is entered in the Регистър на информационните ресурси. We write to those requirements from the first commit rather than discovering them at acceptance.

Most of it is plumbing — and a model only if one is needed at all.

What you are probably thinking

We already paid for a national platform and almost nobody uses it.

Correct, and it is the right objection to open with. Дигитална раница cost around 103 million BGN (about €53 million) under „Образование за утрешния ден“, and the criticism that no usage report was published before МОН committed a further €126,018,597.05 is fair. The answer is not more features. Instrument adoption first — teacher weekly active use, parent monthly active use split by settlement type, resource use per curriculum unit — publish it, and tie every new capability to a measured gap. We agree the adoption metrics before we build, and we show you the ones that look bad.

If AI grades our students, who is liable?

Anything that evaluates learning outcomes or steers the learning process is Annex III point 3(b); proctoring is 3(d) and admission is 3(a). The Digital Omnibus, in force since 27 July 2026, moved most Annex III obligations to 2 December 2027 — that buys time and changes nothing about the design. The exemption route is Art. 6(3), not human review, and it has a hard limit: a system that profiles natural persons is high-risk whatever the oversight around it. Our dropout scoring profiles named students, so we treat it as Annex III 3(b) and 3(a)-adjacent from day one — Art. 9 risk management, Art. 10 data governance, Art. 14 oversight record, Art. 26 deployer obligations, and the Art. 27 fundamental rights impact assessment that applies to you as a public body. Document generation and search do fall outside Annex III under Art. 6(3), and we say which side of the line each component sits on in writing. The architecture still matters — the model drafts, explains, flags and pre-fills; a named human confirms, and the override is logged with who, when and why — but it changes the obligations, not the classification.

Student data cannot leave the country.

Then it will not. BgGPT ships openly in Gemma-2 (2.6B, 9B, 27B) and Gemma-3 (4B, 12B, 27B) with GGUF and GPTQ-W4A16 builds, so Bulgarian-competent inference runs on hardware you control under the Gemma terms. We split the architecture on that line: the model touching identified student data stays on-premise, and offline content generation may use an external API precisely because no personal data reaches it. You get a DPIA, a ROPA entry, a sub-processor register and a data-flow diagram before code — your DPO is a mandatory appointment under GDPR Art. 37(1)(a) and will ask for all four.

Our teachers are near retirement and already overloaded. They will not adopt it.

31.4% of Bulgarian teachers are 55 or older against an EU average of 25.1%, so that is a design constraint rather than a risk to be managed. It means no new login, no new tab and no new mental model: an LTI 1.3 launch from the e-diary they already open every morning, output in the exact document formats the РУО expects, and defaults that are correct for the most common case. We budget training the way МОН did — roughly 13% of the current programme — and we report weekly active use, not seats sold.

You will break our НЕИСПУО reporting, and we cannot miss 20 September.

Which is why nothing we build sits in the critical path on its first cycle. The validation and reconciliation layer runs in shadow for a full year: it tells you what it would have flagged, you compare that against what actually got rejected, and only then does it become the primary path. Capacity is planned around the real peaks — the 15–25 September window, the 1st-to-4th absence campaign, НВО and ДЗИ result days, which run 10–30× median load — not around an annual average. A 99.9% annual uptime figure is meaningless if the outage lands on the 4th.

You are a very small team, and our system fails at 22:00 on the 3rd.

Fair, and the answer is contractual rather than reassuring. Named on-call cover with stated response times across the 1st to the 4th of every month and 15–25 September; a documented runbook your own administrators can execute without us; source code in escrow and, for state work, in the state repository from the first sprint; and a named person on your side trained to re-run the pipeline. If what you need is a 24/7 NOC, we are the wrong supplier and will say so in the ЕЕДОП rather than after award.

We have no budget outside EU programmes, and procurement takes a year.

Then the work is shaped to Programme „Образование“ (ESF+) and NRRP eligibility rules from day one, with deliverables that map to the indicators the grant is measured on. We are used to the fact that the real deadline is a payment claim rather than a release date: documentation retained for the audit trail, costs traceable to eligible activities, and acceptance evidence produced in the form the managing authority asks for. If a capability cannot be made eligible, we say so before it is in the technical specification rather than after the verification.

When we are the wrong choice

Putting a model in the decision path. We do not build systems that rank applicants for admission, award a final mark, or watch students during an exam — Annex III points 3(a), (b) and (d). The honest engineering answer there is a published deterministic rule set with a named human deciding, which is a different product from the one usually being asked for.
Unsupervised answer generation for students. The evidence that AI access cuts study time and proctored retention is strong enough that we will not ship a homework machine and call it learning. Student-facing work has to be built around retrieval, explanation and practice with the teacher in the loop, or we would rather not build it at all.
Replacing НЕИСПУО, your e-diary or your LMS. Those are systems of record with statutory obligations, an eight-vendor market and their own migration programmes. We sit beside them — read their exports, launch over LTI 1.3, write files back in their formats — never instead of them.

Questions we get asked

We are a 180-student school. Can we afford any of this?

Probably not directly, and we would rather say so on the first call. A school that size pays on the order of €310–385 a year (600–750 лв. before the changeover) for its e-diary, partly reimbursed per student through National Programme „ИКТ в системата на предучилищното и училищното образование“, and no bespoke engagement fits inside that. Our buyers are municipalities, РУО, МОН programme teams, publishers and e-diary vendors — organisations serving hundreds or thousands of schools at once. If you are one school with one painful process, ask for the shortest possible answer and we will give it without an invoice.

We run a public procurement. What have you actually delivered in Bulgarian education?

The digital dictionary at beron.mon.bg — contracting authority the Ministry of Education and Science, a public deployment you can open and inspect yourself without asking us for anything. The reference-grade specifics a ЗОП evaluation needs — year, contract or programme reference, scope, and a удостоверение за добро изпълнение if the authority issues one — we supply in writing on request and enter in the ЕЕДОП; we do not print figures on this page that we cannot evidence in a procurement file. Outside education: a European car marketplace handling 300,000+ listings, an NLP module that answers trading questions across email and Instagram with a person approving every reply before it sends — about 85% less manual typing on repeat questions, measured by the client — and email-marketing automation that cut a beauty brand’s outreach time by 60%, also client-measured. That is the complete list. No education awards, no 1EdTech conformance certification yet, no roster of school logos — if the bid requires those, we are the wrong bidder and you should score us accordingly.

We already run Школо. Do we have to replace it?

No, and you should not. Bulgaria has roughly eight competing e-diaries — Школо, АдминПлюс, Сиела, Е-дневник, Феникс and others — plus the НЕИСПУО Дневници module, and the diary is where teachers already live every morning. We read its exports, launch from it over LTI 1.3, and write files back in the formats it and НЕИСПУО expect. Anything that asks a 55-year-old teacher to learn a second daily tool has already failed, whatever the feature list says.

Will our students’ data be used to train a model?

No. Retrieval is not training: approved textbooks and учебни програми are indexed, the model is not fine-tuned on student work, and identified data stays on infrastructure you control. Where fine-tuning genuinely helps, it runs on de-identified or synthetic material with the derivation documented in the DPIA. The ЛОД is retained 50 years under Наредба № 8 as a statutory obligation under GDPR Art. 6(1)(c) — that is a retention duty, not a licence to reuse the record for anything else.

The research says AI makes students learn less. Why would we buy this?

Some of it does, and the strongest study deserves to be taken seriously: a US panel of 3.2 million ALEKS interactions found that study time on AI-susceptible maths problems fell 26.9% cumulatively for college students and 31.3% for high-schoolers; that the time reduction did not appear in proctored practice, where AI was unavailable; and, separately, that performance on proctored retention items — the check that AI use had not been learning — declined 25% in the odds of a correct response. Not a Bulgarian cohort, and we hand over the full reference in writing. That is an argument against unsupervised answer generation for students, not against every application. It is also why our default work targets teacher workload and administrative burden, where the worst failure mode is a wasted draft rather than an unlearned skill.

What do we own at handover, and can we leave?

Assessment content in QTI 3.0, packaged courses in Common Cartridge, rosters and enrolments in OneRoster 1.2, lexical data in TEI Lex-0 with an OntoLex-Lemon projection, learning events as xAPI into an LRS you own, and the pipelines, curriculum URIs and evaluation harness in your repository. The export endpoints are named in the statement of work and demonstrated during acceptance testing, not at termination — and for work for a state authority, the custom source code in the state repository at github.com/governmentbg, in the form ДАЕУ acceptance requires. Handover includes training a named person on your side to re-run and re-score the whole thing without us.

Keep reading

AI advisory and readiness Next-generation AI: retrieval, evaluation, deployment How we work: stages, durations, kill criteria

Book 30 minutes on your September and your 4th of the month

Bring last year’s НЕИСПУО validation report and one month of the absence campaign. On the call we will tell you which failures are a data problem, which are a process problem, and which are one certificate on one machine — and which of the three is worth paying anyone to fix.

Book a call See our projects →

You talk to the engineer who would do the work, and the first two weeks are fixed-fee against your own data.