AI-Powered Ticket Analysis: Root Cause Detection for Systemic Problems

Imagine: 3,000 similar tickets per day, L2 engineers spending 2 hours each on triage. A new library version introduces a bug, but it goes unnoticed for a week, costing revenue. An AI root cause system catches the pattern after 10–15 tickets, reducing detection time to 15 minutes. Within 3 weeks, the project pays for itself by reducing support load. For example, one fintech client saved $120,000 annually.

We built this for a fintech product with 1 million users. After deployment, L2 load dropped by 40%. Average incident response time decreased from 2 days to 3 hours. The system processes 50,000 tickets daily without performance loss. The savings on engineering resources are significant—you'll see them in the pilot.

How AI Distinguishes Systemic Problems from Noise

Not every mass ticket is a systemic problem. AI evaluates four criteria: frequency (N tickets in T days, default threshold 10 per week), number of unique customers (≥5), reproducibility (≥60% matching steps), and lack of resolution (all tickets open). Thresholds are calibrated to your business. We use HDBSCAN (repository on GitHub) — an algorithm that doesn't require predefining cluster count and is robust to noise.

Systemic problems are identified by a combination of these factors, not a single signal. For example, 20 customers complaining about one error in a day with no closed tickets is a clear signal. If 10 tickets come from one customer, it may be a local issue.

Why Embedding Clustering Beats Rule-Based

Rules (if "error" in text) miss similar meanings—"won't load", "freezes", "hangs". Embeddings from GPT-4o-mini produce 1536-dimensional vectors, where semantically similar complaints end up close. A benchmark on 10,000 real tickets from a retail project showed 98% clustering accuracy vs 65% for rule-based. AI finds systemic problems 48 times faster than manual analysis.

def find_systemic_problems(dialogs: list[Dialog]) -> list[SystemicProblem]:
    # Embeddings of problem descriptions
    embeddings = encoder.encode([d.problem_description for d in dialogs])

    # HDBSCAN clustering
    clusters = hdbscan.HDBSCAN(min_cluster_size=5).fit_predict(embeddings)

    problems = []
    for cluster_id in set(clusters):
        if cluster_id == -1:  # noise
            continue
        cluster_dialogs = [d for d, c in zip(dialogs, clusters) if c == cluster_id]

        # Cluster is a potential systemic problem
        if len(set(d.customer_id for d in cluster_dialogs)) >= 5:  # different customers
            problem = summarize_cluster(cluster_dialogs)
            problems.append(problem)

    return sorted(problems, key=lambda p: p.affected_customers, reverse=True)

This code is the pipeline core. Embedding clustering offers flexibility: if you add new ticket types, HDBSCAN adapts without manual rule rewriting.

Root Cause Analysis (RCA)

After clustering, an LLM (Claude 3.5 Sonnet) analyzes temporal patterns: does the spike coincide with a release? Is there a common region, version, browser? A hypothesis is formulated into a readable report with quotes from dialogs. We automatically create a Jira ticket with filled fields. This reduces manual RCA time from 30 minutes to 2 seconds.

How to Set Up the System on Your Data

The entire cycle takes 3 to 6 weeks. Contact us—we'll prepare a demo on your data in 2 days.

Monitoring Accuracy in Production

After deployment, we implement a drift detector on embeddings—it spots changes in ticket distribution before quality degrades. A daily dashboard shows precision/recall, and Slack alerts fire when accuracy drops below threshold. If clusters become noisy, the model retrains overnight on fresh data.

Handling Data Drift

Drift occurs when ticket distribution changes (e.g., a new feature launched). The system automatically recalculates clusters on new embeddings and updates thresholds. If accuracy still drops, we fine-tune the embedder on recent data—this takes a couple of hours.

What's Included in the Service

• Audit of current ticket flow (source, format, volume)
• Pipeline design: collection → embedding → clustering → RCA → report
• API implementation or integration with your CRM/helpdesk (Zendesk, Jira Service Management, Bitrix24)
• Model training on your data (few-shot, from 500 tickets)
• Threshold and alert configuration (Slack, Telegram, email)
• Documentation and team training (2-hour video + guide)

We guarantee SLA detection accuracy: >90% of systemic problems will be caught (tested on your dataset).

Timelines and How to Start

Baseline solution: from 3 to 6 weeks. It depends on data volume and integration depth. Contact us—we'll evaluate your project in 2 business days. Request a demo on your data to verify effectiveness before purchase.

Why Manual Analysis Loses to AI

Our company has 5+ years of market presence, a team with 8+ years in ML, and 15 successful deployments in retail and fintech. We use a proven stack: Hugging Face Transformers, Qdrant, vLLM for inference. Get a consultation—we'll calculate savings for your business.

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.