AI Classification and Routing System for Citizen Appeals

Manual Appeal Routing: Lost Time and Errors

In government agencies, up to 30% of initial appeals reach the wrong executor. The applicant receives a formal refusal "not in our competence," while the responsible employee spends time on redirection. Result — missed review deadlines and increased repeated complaints. We developed an AI classifier that solves this: Accuracy@1 reaches 92–95% immediately after adaptation. Get an expert consultation — we will analyze your current routing scheme.

How Does Classification Work and Why Is It More Accurate Than Manual Distribution?

Government rubrics are hierarchical: federal → departmental → territorial. An appeal must be assigned to the correct level and specific executor. We use a combination of semantic search on precedents and LLM parsing. Manual distribution gives Accuracy@1 of 60–70% and 20–30% false redirections. The AI classifier reduces false routing rate to 5% by considering territorial affiliation, unstructured attachments, and temporal context.

class RequestClassification(BaseModel):
    federal_rubric: str
    department_rubric: str
    responsible_unit: str
    responsible_officer: str | None
    territorial_scope: str
    subject_area: str
    requires_field_inspection: bool
    other_agencies: list[str]
    confidence: float

def classify_request(text: str, attachments: list) -> RequestClassification:
    # Semantic search on precedent database
    precedents = precedent_db.search(text, top_k=10)
    # Multimodal classification if attachments exist (photos, documents)
    if attachments:
        attachment_context = analyze_attachments(attachments)
    else:
        attachment_context = ""
    return llm.parse(
        build_routing_prompt(text, attachment_context, precedents),
        response_format=RequestClassification
    )

Special Cases: Attachments, Geolocation, Mixed Appeals

Attachments (photos, scans, PDFs) often contain key information. The system automatically extracts text, recognizes objects (e.g., potholes in photos), and geotags. If an attachment is unreadable, the model relies on the appeal text and marks the case as requiring expert assessment.

def extract_and_geolocate(text: str) -> GeoContext:
    addresses = ner_model.extract(text, entity_type="ADDRESS")
    located_objects = []
    for addr in addresses:
        coords = geocoder.geocode(addr)
        if coords:
            admin_unit = geodata.get_admin_unit(coords)
            responsible = routing_matrix.get_responsible(
                issue_category=...,
                admin_unit=admin_unit
            )
            located_objects.append(GeoObject(
                address=addr,
                coords=coords,
                responsible_org=responsible
            ))
    return GeoContext(objects=located_objects)

Geocoding — via Yandex Geocoder or Nominatim. Accuracy of municipality determination — 95%+.

One appeal often contains several diverse issues (e.g., "noisy neighbors and broken elevator"). The system splits the text into independent segments, classifies each separately, then consolidates responses. The review time is calculated based on the longest segment.

Classifier Training and Metrics

Base — historical data. Critical stage — cleaning from erroneous redirections (exclude appeals that were forwarded). Supplemented with anonymized data from similar agencies. We use Active Learning: experts label complex cases, model is fine-tuned iteratively. For quality assessment on new rubrics, we use few-shot testing.

What Does Automation Provide?

Through automation, employee workload is reduced — they stop being "sorters" and focus on substantive work. Time savings: from 15 minutes per appeal. For an agency with a flow of 1000 appeals per day, this is more than 200 working hours per month, equivalent to savings from 1.2 million rubles per year.

Integration with Document Management Systems

We support EDMS: DELO, DIRECTUM, Docsvision, 1C:Document Management. The system transfers the appeal with a filled registration card — the executor receives not an email but a structured task with a set deadline.

Process and Implementation Timeline

1. Audit of current routing process and data.
2. Development of the classification model (fine-tuning LLaMA 3 or GPT-4o, ensemble with BERT classifier).
3. Creation of a vector precedent database (Qdrant, 1536-dim embeddings).
4. Integration with your EDMS (API setup, cards, statuses).
5. Analytics and monitoring dashboard.
6. Documentation, operator training, 6 months warranty support.

Pilot project — from 4 weeks. Full implementation with integration — from 3 months. Average pilot project budget — from 300 to 500 thousand rubles, full implementation — from 1 to 3 million rubles. Exact timelines and cost are determined after auditing your system.

Typical Errors and Their Handling

• Appeals with incomplete data (no address or subject).
• Rubric synonymy (different names for the same topic).
• Attachments without text (only images) — the model marks them as "requires expert review."

We provide handling for each of these cases. We guarantee Accuracy@1 not lower than 90% after adaptation, otherwise we will refine the model for free.

What Is Included in the Work?

• Full audit of current routing scheme and data.
• Development and customization of the classification model.
• Integration with your EDMS (API, cards, statuses).
• Vector precedent database with embeddings.
• Analytics and monitoring dashboard.
• Operator training and technical documentation.
• 6 months warranty support.

Contact us for a free audit. Order a pilot project — evaluate the result on your own data.

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.