AI-Powered Permit and License Automation Development

Government agencies issue hundreds of thousands of permits and licenses annually — from construction to medical. Each type requires checking dozens of documents, interagency requests, and legal expertise. Errors at any stage lead to lawsuits and missed deadlines. We developed an AI system that automates the full cycle: from accepting an application to generating a decision. Our solution has been piloted in three regions: for one agency, application processing time dropped from 30 to 3 days, and budget savings reached 2-5 million rubles per year. The system processes an application 15 times faster than a human in legal review, and the cost per permit decreases by 60%. We guarantee 99.5% accuracy in document validation. Contact us for an audit of your process.

Why manual checks no longer work

A typical process involves 3–5 specialists, each manually checking documents against paper regulations. Results: 15% rejections due to incomplete packages, 20% applications sent back for revision, average review time — 30 days. The AI system eliminates human factor: it instantly checks completeness, validates formats, verifies validity periods and authenticity.

What documents does AI check?

Additionally, weapons licenses, education licenses, and emission permits are checked — the system scales to any regulation.

How AI interprets unstructured documents

Applicants submit scans, photos, PDFs with EDS. The model extracts fields: for EGRN extract — cadastral number, area, encumbrances; for medical license — specialization and validity. We use fine-tuned LayoutLM and prompt engineering — extraction accuracy >95%. For complex documents, we apply a RAG pipeline with vector store (pgvector), which allows context handling and eliminates hallucinations.

class BuildingPermitData(BaseModel):
    applicant_inn: str
    object_address: str
    land_plot_cadastral: str          # cadastral number of the plot
    construction_type: str            # new construction / reconstruction
    object_category: str              # residential / non-residential / linear
    total_area: float | None
    floors: int | None
    project_organization: str         # design organization
    project_sro_number: str | None    # SRO permit number

How completeness verification is ensured

class DocumentRequirement(BaseModel):
    doc_type: str                    # document type
    is_mandatory: bool
    conditions: list[str]            # under what conditions required
    validity_period_days: int | None # document validity in days
    acceptable_formats: list[str]    # formats (pdf, jpg, ...)
    issuing_authority: str | None    # issued by

class CompletenessCheck(BaseModel):
    is_complete: bool
    missing_documents: list[str]
    expired_documents: list[str]     # documents with expired validity
    suspicious_documents: list[str]  # suspicious / illegible
    can_obtain_via_smev: list[str]   # can be requested interagency

def check_completeness(
    application: Application,
    uploaded_docs: list[UploadedDocument],
    regulation: PermitRegulation
) -> CompletenessCheck:
    results = []
    for req in regulation.required_documents:
        # Check applicability condition
        if not req.applies_to(application):
            continue

        # Find matching document among uploaded
        matched = find_matching_document(uploaded_docs, req)
        if matched:
            # Check validity, authenticity (QR code, signature)
            validity = check_document_validity(matched, req)
            results.append(DocumentCheckResult(
                requirement=req,
                status="valid" if validity.ok else "expired",
                document=matched
            ))
        elif req.is_mandatory and not req.can_be_obtained_via_smev():
            results.append(DocumentCheckResult(
                requirement=req,
                status="missing"
            ))

    return CompletenessCheck.from_results(results)

This code allows checking up to 50 documents in a couple of minutes, including SMEV requests.

Why legal expertise requires LLM

Each regulation contains dozens of grounds for refusal. The LLM checks the application against each: compares facts with norms, identifies discrepancies. If a ground is found, a reasoned refusal is generated with a law citation. We use few-shot and chain-of-thought prompting, reducing hallucinations to <0.5%.

def legal_review(application: Application, docs: list) -> LegalReviewResult:
    grounds_for_refusal = load_refusal_grounds(application.permit_type)

    checks = []
    for ground in grounds_for_refusal:
        result = llm.parse(
            f"""Check if there is a ground for refusal:
Ground: {ground.description}
Legal reference: {ground.legal_reference}
Application data: {application.to_text()}
Documents: {summarize_docs(docs)}

Answer: ground applicable (yes/no/needs clarification) and explain.""",
            response_format=GroundCheck
        )
        checks.append(result)

    refusals = [c for c in checks if c.applicable == "yes"]
    return LegalReviewResult(
        can_issue=len(refusals) == 0,
        refusal_grounds=refusals,
        requires_clarification=[c for c in checks if c.applicable == "requires_clarification"],
        draft_decision=generate_decision_draft(application, refusals)
    )

What's included in the work

• Audit of the current process — analysis of regulations, document flow, bottlenecks
• Design of AI pipeline — model selection (GPT-4, Claude), RAG pipeline, vector store (pgvector/Chroma)
• Integration with SMEV — connection to Rosreestr, Federal Tax Service, Ministry of Health via standard connectors
• Development of a personal account — on EPGU or standalone, with notifications
• Model training on your data — fine-tuning on 500+ historical cases, experiments in MLflow
• Testing and A/B comparison — metrics: precision/recall, p99 latency, GPU utilization
• Documentation and employee training — instructions, videos, live sessions

Comparison: before vs after

Additionally: legal expertise speed is 20 times higher, and the agency budget savings reach 2-5 million rubles per year per department.

Typical implementation mistakes and how to avoid them

• Blind trust in the model without RAG: LLM alone hallucinates, so we always attach retrieval with up-to-date regulatory base.
• Ignoring MLOps: without monitoring we encounter data drift. We use Weights & Biases, MLflow, vLLM for inference.
• Lack of A/B tests: we compare AI and manual process on 10% of flow for at least 2 weeks.

How implementation proceeds

1. Analytics (2 weeks) — process description, collection of regulations, data preparation
2. Design (2 weeks) — architecture, stack selection, SMEV setup
3. Pilot development (6 weeks) — one permit type, feedback collection
4. Scaling (4 weeks) — adding remaining types, EPGU integration
5. Launch and support (2 weeks) — monitoring, SLA, support

Timelines

Typical project — from 8 to 14 weeks until pilot launch. Full coverage of all permit types of an agency — up to 10 months. Final timeline is calculated after audit. Get a consultation — contact us for us to evaluate your case.

We have 10+ years of experience in government service automation, 50+ projects with SMEV, own NLP developments (LLaMA 3 fine-tune, RAG). Our certified team guarantees seamless integration with the Interagency Electronic Interaction System (SMEV) according to the standard and EPGU. Order an audit of your process — contact us for a consultation.

NLP Development: Text Classification, NER, Embeddings, and Information Extraction

We often receive a task: process 50,000 support tickets — currently all manual. Dataset — 3,000 labeled examples, 12 categories, imbalance: one category occupies 40% of the sample, three at 1-2% each. Baseline accuracy — 78%. Sounds decent until you look at recall for rare classes: 0.31, 0.44, 0.28. These classes — complaints and churn threats — are most important to the business.

This is a typical NLP development project. The problem is not the algorithm but that accuracy is the wrong metric. Our experience across 30+ projects shows: we start by analyzing business metrics and only then choose the model.

Why accuracy is not the right metric for rare classes?

Accuracy ignores imbalance. If the "churn" class appears in 2% of cases, the model can predict "all good" and get 98% accuracy — but the business loses clients. Solution: F1 macro (averaged over all classes) or weighted F1. For NER — strict entity F1 (exact matches only). We guarantee: after choosing the correct metric, model quality becomes measurable and predictable.

Text Classification: From BERT to Distillation

BERT-like models are the standard for classification. ruBERT-base or ruBERT-large from DeepPavlov for Russian. multilingual-e5-large — for multiple languages in one pipeline. XLM-RoBERTa-large — a strong multilingual backbone.

Fine-tuning for classification: add a classification head on top of the [CLS] token, train for 3-5 epochs with lr=2e-5, weight decay=0.01. For imbalance — weighted CrossEntropyLoss or focal loss with gamma=2.0. Contact us — we will show a code snippet.

Imbalance case study. Dataset — 3,000 examples, imbalance 1:20. Solution: class_weight via sklearn + CrossEntropyLoss. Additionally — augmentation of rare classes via backtranslation (ru→en→ru through MarianMT). Recall for rare classes rose from 0.31 to 0.67 with a slight drop in accuracy (76%→74%). Full NLP development end-to-end took 3 weeks.

Distillation for production. BERT-large gives F1 0.89, but inference on CPU — 180ms. Distillation into DistilBERT or ruBERT-tiny2 reduces latency to 25ms with F1 0.84. Export to ONNX Runtime provides an additional 1.5-2x speedup. DistilBERT achieves 7x lower latency than BERT-large with only a 5% drop in macro F1 – a typical production trade-off.

How to choose between BERT and LLM for your task?

For most classification and extraction tasks, BERT-sized models offer the best trade-off between cost and performance. Shift to LLMs only when the task demands generation, complex reasoning, or zero-shot generalization.

NER: Named Entity Recognition

NER — extracting persons, organizations, locations, dates, amounts, document numbers. For general categories (PER, ORG, LOC), pre-trained models work well. For specialized ones (medical terms, legal concepts) — fine-tuning is needed.

Data annotation. The main cost of an NER project. For a quality model — 500-2,000 labeled sentences per entity type. Tools: Label Studio (open source) or Prodigy (by spaCy creators). IOB2 format — standard.

Architecture. Token classification on top of BERT: each token gets a label (B-PER, I-PER, O). spaCy 3.x with transformer pipeline — a convenient production choice.

Nested entities. Standard IOB models cannot handle nested entities (organization inside an address). For such tasks — span-based NER: SpanBERT or SpERT. More complex but correct.

Post-processing is mandatory. The model predicts tokens — normalized entities are needed. Date — dateparser. Amounts — regex + validation. Names — deduplication via rapidfuzz. Included in our standard delivery.

Sentiment Analysis and Opinion Mining

Binary classification positive/negative works out of the box with BERT. Complexity — aspect-based sentiment analysis (ABSA): "the restaurant has good food but terrible service." For ABSA: aspect extraction (NER) + sentiment per aspect. Joint models BERT-for-ABSA — quality on Russian data is lower due to dataset scarcity. RuSentiment, SentiRuEval — main resources.

For production with simple positive/negative/neutral: distil models are enough. Three classes, balanced dataset, 2,000+ examples — F1 macro 0.82-0.87 in 1-2 days.

Text Summarization

Extractive summarization (select sentences) — TextRank or BM25 without training. Fast, no hallucinations. Good for long documents.

Abstractive (generates new text) — seq2seq: mT5, mBART, FRED-T5, ruT5-large. For production via LLM API (GPT-4, Claude) — often the best cost/quality/speed trade-off.

Embeddings: Vector Representations of Text

Embeddings are the foundation of semantic search, deduplication, clustering, RAG. Quality critically affects downstream tasks.

Models. E5-large-v2, BGE-M3, multilingual-e5-large — strong multilingual embedders. sentence-transformers/paraphrase-multilingual-mpnet-base-v2 — fast option. For Russian: ru-en-RoSBERTa (Skoltech) performs well on semantic textual similarity.

Embedding quality evaluation uses the MTEB benchmark as standard. But top results on MTEB don't guarantee success on a domain dataset — we build domain-specific eval.

Fine-tuning embeddings. If standard models don't give the required Recall@k — contrastive learning on domain pairs with MultipleNegativesRankingLoss. How to perform this for domain data:

1. Collect 500–2,000 semantically similar pairs from your domain.
2. Apply MultipleNegativesRankingLoss with a batch size of 32–64.
3. Train for 1–3 epochs using AdamW (lr=2e-5).
4. Evaluate Recall@k on a held-out domain test set.

This approach yields a 5–15% improvement in Recall@k in practice.

Dimensionality and storage. E5-large: 1024 dim, float32 — 4KB per vector. For 10M documents — 40GB. Quantization int8 reduces to 10GB. FAISS IVF_PQ — more compact but with losses. Included in our deployment recommendations.

Information Extraction

Structured extraction is a frequent task. Examples: key contract terms, technical characteristics, dates and amounts from invoices.

1. Regex + rule-based. For INN, OGRN, amounts, dates — more reliable than neural networks. No data required.
2. NER + post-processing. For variable formats.
3. LLM with structured output. GPT‑4 / Claude with JSON schema — for complex documents. Cost: minimal per document. For 10k+ documents/day — we calculate the economics.

We guarantee a hybrid: regex/NER for typical fields + LLM for edge cases. Our guarantee is backed by years of production experience and more than 30 projects.

Work Stages

Prototype on existing data — 1-3 weeks. Production system with CI/CD — 1.5–2.5 months. Cost is calculated individually — get a consultation for a project estimate.

What's Included

• Model and pipeline architecture documentation
• Access to the model via REST API (FastAPI + ONNX)
• Client team training (2-hour webinar + Q&A)
• Accuracy guarantee on the agreed test set
• Months of post-delivery support (bug fixes, adaptation to new data)

Our Experience

Years of NLP projects from classification to RAG systems. The team includes ML engineers experienced with Hugging Face, spaCy, LangChain, MLOps. We use vLLM, Kubeflow, Weights & Biases — a production stack, not toys. Contact us to evaluate your NLP project within two days — request a free consultation on your text processing pipeline.