LLM Pipeline for Automatic Meeting Summarization

Automatic Summarization of Meeting Transcripts: An LLM Pipeline

A 60-minute meeting generates 10,000–12,000 words of transcript. 80% of that volume is context, repetitions, and conversational filler. Without a summarization system, you spend 15–30 minutes manually extracting decisions — and the result suffers from subjectivity and omissions. We implement automatic summarization pipelines using LLMs that extract the semantic core in seconds and produce a structured summary with topics, decisions, and next steps.

In one project, we processed meetings for a team of 20 people — 3–5 meetings per day. After deployment, the time to prepare a summary dropped from 30 minutes to 20 seconds per meeting, and the quality became consistent. Time savings per meeting: up to 30 minutes, which at 50 meetings per month translates to 25 hours. Typical time savings: up to 95% per meeting.

How the Summarization Pipeline Works

The pipeline takes raw transcript (plain text or JSON with speaker labels) and returns a summary in this format:

• Short summary (2–3 sentences)
• Key topics
• Decisions made
• Open questions
• Participants and their positions

For short meetings (under 30 minutes, <6000 tokens), we use a direct prompt — latency 5–15 seconds. For long meetings, we apply map-reduce:

[Transcript]
    → [Preprocessing: split into chunks of 3000 tokens]
    → [Map: summarize each chunk]
    → [Reduce: synthesize final summary]
    → [Structuring: topics, decisions, next steps]

Comparison: Manual vs. Automated Summarization

An LLM pipeline summarizes a meeting 30 times faster than a human and eliminates subjectivity.

Which LLMs to Choose for Summarization?

Model selection depends on speed, cost, and quality requirements. For everyday meetings (status updates, planning), compact models like GPT-4o-mini or LLaMA 3 8B are optimal — they process short transcripts in 5–10 seconds and cost pennies. For technical discussions (architecture, code reviews), we use GPT-4o or Mistral Large — their deep context understanding reduces hallucinations. If data is confidential, we deploy locally with Qwen 72B or Mistral 7B using INT8 quantization. In any case, we use few-shot prompts with 2–3 examples to stabilize the output format.

Benefits of Automating Summarization

Manual summarization takes 15–30 minutes per meeting and is subjective. An LLM pipeline delivers a structured result in 10–60 seconds, consistently across all meetings. Our engineers, with 5+ years of experience, guarantee quality at the level of a senior analyst. For most companies, the processing cost per meeting is negligible, and the time savings pay off within the first few weeks. Contact us for a preliminary assessment — we will design an architecture for your volumes and demonstrate a working pipeline. Request a consultation to discuss your scenarios.

What's Included in the Work?

1. Audit of current processes: transcript format, sources, summary requirements.
2. Pipeline design: model selection, chunk definition, prompt engineering with few-shot examples.
3. Integration with sources: Zoom (Whisper + API), Google Meet (Speech-to-Text), Microsoft Teams (Graph API), Fireflies.ai / Otter.ai (webhook).
4. Deployment: containerization (Docker), deploy to your infrastructure or cloud.
5. Documentation and team training.
6. Support for 30 days after launch.

from langchain.chains.summarize import load_summarize_chain
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.chat_models import ChatOpenAI

llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)
text_splitter = RecursiveCharacterTextSplitter(chunk_size=3000, chunk_overlap=200)
chain = load_summarize_chain(llm, chain_type="map_reduce")

Example output for a "Sprint Review" meeting:

{
  "summary": "The team discussed progress on three tasks. Decision: extend deadline for task A by 2 days.",
  "topics": ["Progress on task A", "Blocker on task B", "Sprint planning"],
  "decisions": ["Task A deadline moved to Friday", "Decompose task B into subtasks"],
  "action_items": ["@alice: update Jira", "@bob: estimate effort"]
}

Integrating the Pipeline with Your Sources

• Zoom — Zoom AI Companion API or Download recordings API + Whisper for transcription.
• Google Meet — Google Meet API + Speech-to-Text.
• Microsoft Teams — Graph API transcripts.
• Fireflies.ai / Otter.ai — webhook with ready transcript.

The result is saved to Notion, Confluence, Jira, or corporate wiki via their APIs. The approach is described in the LangChain documentation.

Timeline and Cost

Timeline: from 5 to 20 days depending on integration complexity and the need for map-reduce. Cost is calculated individually — depends on the number of sources, required accuracy, and custom prompts. Get a consultation: our engineers will analyze your processes and propose the optimal solution.

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.