AI-Generated Structured Meeting Minutes from Transcripts

You spent an hour in a meeting, but the secretary takes half a day to transcribe and format the minutes. Or worse — the minutes don't reflect actual decisions, and a month later nobody remembers what was approved. In large companies with dozens of meetings daily, such delays lead to missed deadlines and up to 20% of managers' time wasted on clarifications. We solve this: AI automatically generates legally valid minutes from raw transcription, saving 70% of lawyers' and secretaries' time. Each minute undergoes validation via chain-of-thought prompting and RAG check against previous records, eliminating fabricated facts. Our experience — 5+ years in MLOps, over 50 document workflow automation projects. The cost of automation pays off in 3–4 months due to reduced man-hours.

AI-Generated Meeting Minutes: From Transcription to Document

The process consists of three phases. Phase 1 — Metadata extraction: Whisper transcribes audio, LLM extracts date/time, participants with positions, agenda. Phase 2 — Content structuring: each agenda point → discussion → decision/voting/deferred. LLM processes sections sequentially with chain-of-thought to minimize hallucinations. Phase 3 — Formatting into template: python-docx inserts data into DOCX template bookmarks. The final document is sent to participants for confirmation. The AI method is 3 times faster than manual with comparable quality.

Why Minutes Are a Legally Binding Document

Meeting minutes are a legally binding document: they contain date, participant list, agenda, decisions, votes, and signatures. AI generation must exclude fabricated facts (hallucination). We use few-shot prompts with real minute examples and post-processing: checking consistency of decisions with the agenda. According to corporate law practice, lack of signatures may reduce the legal force of minutes.

Typical Problems and Their Solution — AI Generation of Structured Minutes

• Incorrect name recognition: use contextual correction from the organization's contact database.
• Different date formats: the template contains a mask per corporate standard.
• Missed action items: LLM additionally scans each utterance for assignments with deadlines.

How to Minimize Hallucinations in Minutes?

The main challenge of generation is fabricated facts. We apply chain-of-thought prompting: each agenda item is processed separately with stepwise reasoning. Additionally, we use RAG (Retrieval-Augmented Generation) with ChromaDB: previous minutes of this meeting are loaded into context to maintain consistency. P99 generation latency — 3 seconds at 8K token context size.

Comparison of Approaches

Another table for accuracy comparison:

Our Stack and Experience

Stack: OpenAI GPT-4 (structuring), LlamaIndex (context reordering), ChromaDB (meeting storage for RAG), python-docx (generation). To reduce latency we use INT8 quantization via vLLM — p99 latency < 3 sec.

In one project for a consulting company with 500 meetings per month, we reduced minutes preparation time from 3 hours to 12 minutes, and revision returns dropped by 90%. Document workflow budget savings reach 80%. Order a pilot for 3–5 days — see the savings yourself. Get a consultation on your project.

Process of Work

1. Analytics — study corporate templates, audio sources (Zoom, Teams, files), storage requirements.
2. Design — create a DOCX template with bookmarks, configure a metadata schema.
3. Implementation — set up pipeline: transcription → extraction → structuring → generation.
4. Testing — run on 50 real recordings, adjust prompts based on logs.
5. Deployment — deploy in infrastructure (on-prem/cloud), connect API.
6. Maintenance — quality monitoring, prompt updates, support.

Timeline and What's Included

Timeline — from 1 day to 2 weeks (depending on template complexity and integrations). What's included in the work:

• API and architecture documentation
• Operator training
• 24/7 support for the first 30 days
• Stability guarantee (99.9% availability)

Contact us to discuss your project details and get a consultation. Find out how AI-generated minutes can save your budget this quarter.

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.