Prospekt

Compliance API for Russian-language text. Cyrillic spelling — Проспект.

What the service does

Prospekt answers a single question — “Can this text be published under Russian law?” — and answers it in the way platforms and AI services need: one HTTP request, structured JSON, immutable audit log on by default.

The service covers 24 articles of the Russian Criminal Code (УК РФ), the Code of Administrative Offences (КоАП РФ), and adjacent federal laws: discrediting the Armed Forces of the Russian Federation (art. 207.3 УК РФ), incitement to extremism (art. 280 УК РФ), rehabilitation of Nazism (art. 354.1 УК РФ), distribution of misleading public-interest information (art. 13.15 КоАП РФ), promotion of non-traditional sexual relationships (art. 6.21 КоАП РФ), foreign-agent labelling (255-ФЗ), personal data protection (152-ФЗ), the children protection statute (436-ФЗ), and others. The full taxonomy is in the project documentation.

Under the hood — a five-layer cascade: dictionaries and regular expressions; a small fine-tuned classifier (1—3 B parameters); a large language model running in a Russian compute perimeter, augmented with a corpus of court rulings; conformal-prediction calibration of aggregated scores; a manual review queue for contentious cases. Recall by category group ranges from 70—80 % on defamation to 99.9 % on lexical patterns such as direct leakage of personal data. The numbers are open and reproducible: the 1B model is freely available on Hugging Face under the Apache 2 licence.

Who it is for

Prospekt is designed around three scenarios.

AI-first startups and LLM products. Pipe generated text through /v1/classify before showing it to the user. The response carries per-category scores and an audit_id for any future complaint review. Pro and Business tiers are self-serve via card payments through ЮKassa.

UGC platforms. Pre-filter before the human moderation queue: 95—97 % of inputs pass automatically as safely cleared, 3—5 % go to a moderator. On a flow of 500 000 items per day this frees 25—30 moderator positions — roughly ₽22—27 million per year.

Corporate compliance. Banks, media, and large e-commerce — install in the customer’s perimeter with domain-specific fine-tuning and a dedicated account manager. ФСТЭК certification is on the roadmap for Q3 2027.

What runs the service

• Response time: p95 ≤ 250 ms, p99 ≤ 500 ms for texts up to 2 000 characters.
• Models: open 1B model on Hugging Face under Apache 2, and a production 7B model running on servers in Russia (Pro tier and above).
• Audit log: text hash, scores, model version, timestamp. Retention — one year, in line with the requirements for organisers of information distribution (374-ФЗ, RF Government Decree № 728).
• Client-pinned model version — pinnable for up to a year, with 90 days of advance notice on every release.
• Per-category deploy gate: a new model version reaches production only if it is no worse than the previous one on every individual category — averaging is forbidden, since it can mask regressions on narrow but legally consequential categories.

What Prospekt is not

The service does not block content, does not forward text to third parties, and does not notify regulators. It is not a replacement for the editor, the lawyer, or the moderation team. Decisions to publish, remove, or escalate are made by the client based on the returned scores and the client’s own thresholds.

Materials

• Landing page: ai-moderator.dev.borisovai.ru
• Sales deck: ai-moderator.dev.borisovai.ru/pitch.html — 15 slides on the product, the market, pricing, the rollout plan, and common buyer questions.
• API console prototype: ai-moderator.dev.borisovai.ru/prototype.html — workspace with /v1/classify examples, the audit log, category thresholds, and team settings.
• Repository: gitlab.dev.borisovai.ru/soft/limes

Status

Closed beta. The open 1B model is already published on Hugging Face — download it and run it on your own texts. The production 7B model and the Pro / Business tiers go live in Q3 2026. ФСТЭК certification for the corporate tier is targeted for Q3 2027.

Requests for early access and pilot deployments — at info@borisovai.ru.