Multimodal AI Pipeline: Text, Image, and Audio Processing

Imagine you need to process an incoming document—a scanner outputs a PDF with an invoice, a doctor dictates an audio commentary, and a client sends a photo of a defective part. Each data type requires its own tool, and gluing them into a single multimodal AI pipeline is the task of a multimodal AI system. We implement such solutions for manufacturing, medicine, and logistics: this reduces manual work by 60–80% and lowers errors to 0.5% according to our project data. The core stack is a Vision LLM (e.g., Claude Sonnet 4.5), Whisper: Robust Speech Recognition via Large-Scale Weak Supervision (Whisper large-v3) for audio, and a custom TTS model. Typical latency: 2–5 seconds per complex request.

Key challenges addressed by multimodal pipelines

Heterogeneous data formats. A client sends photos, scans, audio recordings, videos—all need to be combined and analyzed in a single context. Without a multimodal pipeline, you end up using separate utilities, leading to context loss and errors. For video, we extract key frames, analyze motion and defects.

Recognition quality. Old OCR systems fail on non-standard fonts, Whisper struggles with accents and noise. We select model combinations, fine-tune for your domain, and add post-processing with LLM validation.

Latency. Sequential processing through multiple APIs gives latency >30 seconds. We build parallel pipelines with queues (RabbitMQ/Kafka) and embedding caching—typical response time is 2–5 seconds.

Multimodal fusion architecture

The basic architecture includes four steps:

1. Audio → text via Whisper (model whisper-1, supports 99 languages).
2. Images—encoded as base64 and passed to a Vision LLM (Claude, GPT-4o). For large images, we apply adaptive resizing while preserving quality.
3. Documents—extract text via pdfplumber / PyMuPDF, then merge with text.
4. Fusion—all modalities are collected into a single content array and sent to the LLM with the task prompt.

Below is a Python (asyncio) code example we use in production:

from anthropic import AsyncAnthropic
from openai import AsyncOpenAI
import base64
import aiohttp
from pathlib import Path
from typing import Union
from pydantic import BaseModel

anthropic_client = AsyncAnthropic()
openai_client = AsyncOpenAI()

class MultimodalInput(BaseModel):
    text: str | None = None
    image_paths: list[str] = []
    audio_path: str | None = None
    document_path: str | None = None

class MultimodalPipeline:

    async def process(self, input_data: MultimodalInput, task: str) -> str:
        """Process multimodal input"""
        content_blocks = []

        # 1. Audio → text
        if input_data.audio_path:
            transcript = await self.transcribe_audio(input_data.audio_path)
            content_blocks.append({"type": "text", "text": f"[Audio transcription]: {transcript}"})

        # 2. Images
        for image_path in input_data.image_paths:
            image_block = await self.prepare_image(image_path)
            content_blocks.append(image_block)

        # 3. Documents (PDF)
        if input_data.document_path:
            doc_text = await self.extract_document(input_data.document_path)
            content_blocks.append({"type": "text", "text": f"[Document]:\n{doc_text}"})

        # 4. Text query
        if input_data.text:
            content_blocks.append({"type": "text", "text": input_data.text})

        content_blocks.append({"type": "text", "text": f"\nTask: {task}"})

        response = await anthropic_client.messages.create(
            model="claude-sonnet-4-5",
            max_tokens=4096,
            messages=[{"role": "user", "content": content_blocks}],
        )

        return response.content[0].text

    async def transcribe_audio(self, audio_path: str) -> str:
        """Transcribe audio via Whisper"""
        with open(audio_path, "rb") as f:
            transcription = await openai_client.audio.transcriptions.create(
                file=(Path(audio_path).name, f.read()),
                model="whisper-1",
                language="ru",
            )
        return transcription.text

    async def prepare_image(self, image_path: str) -> dict:
        """Prepare image for Claude Vision"""
        if image_path.startswith("http"):
            async with aiohttp.ClientSession() as session:
                async with session.get(image_path) as resp:
                    image_data = base64.b64encode(await resp.read()).decode()
            media_type = resp.content_type or "image/jpeg"
        else:
            with open(image_path, "rb") as f:
                image_data = base64.b64encode(f.read()).decode()
            ext = Path(image_path).suffix.lower()
            media_type = {"jpg": "image/jpeg", "jpeg": "image/jpeg",
                         "png": "image/png", "gif": "image/gif",
                         "webp": "image/webp"}.get(ext[1:], "image/jpeg")

        return {
            "type": "image",
            "source": {"type": "base64", "media_type": media_type, "data": image_data}
        }

    async def extract_document(self, pdf_path: str) -> str:
        """Extract text from PDF document"""
        import pdfplumber

        text_parts = []
        with pdfplumber.open(pdf_path) as pdf:
            for page in pdf.pages:
                text = page.extract_text()
                if text:
                    text_parts.append(text)

        return "\n\n".join(text_parts)[:10000]

Practical case: invoice processing

One of our clients—a logistics company—received 500+ invoices daily. Manual entry took 15 minutes per document and had an 8% error rate. We deployed a multimodal pipeline:

• Scan of invoice → Claude Vision extracts table, amounts, signatures.
• If there is an audio comment (e.g., "correct to 2000 rubles") — Whisper transcribes, LLM corrects the JSON.
• Result is written to 1C via API.

Results: processing time dropped from 15 minutes to 2 seconds, errors from 8% to 0.5%. Annual savings of $60,000 (Client case study).

How we build a multimodal pipeline

1. Task audit—you send sample data, we measure baseline and set goals.
2. Design—select models: Claude Sonnet for vision, Whisper large-v3 for audio, TTS-1 for synthesis. Optimize pipeline.
3. Implementation—write code (Python, async), set up queues, cache, monitoring.
4. Testing—A/B test against old system, record improvement.
5. Deployment—containerize, deploy in your environment (on-prem or cloud).

Why trust us with a multimodal pipeline?

With 5+ years of AI experience and 50+ implemented projects, we build production-grade multimodal AI pipelines. We use production stacks: PyTorch, Hugging Face, LangChain, vLLM. We guarantee quality through monitoring p99 latency and accuracy on your data. Certified specialists in Claude, GPT-4, and Whisper. We work with any models: from open-source LLaMA to proprietary APIs. For fine-tuning we use LoRA and QLoRA on your data. We always offer the optimal balance of speed and accuracy.

For searching multimodal data we use RAG architecture with vector DBs (ChromaDB, Qdrant). This allows fast retrieval of relevant information from large archives of documents, images, and audio recordings, improving query accuracy.

Estimated timelines and budget

Cost is calculated individually—depends on number of models, need for fine-tuning, and infrastructure. Contact us for an accurate estimate.

What's included

• Pipeline source code with documentation.
• Models and weights (if fine-tuning is used).
• Integration with your systems (API, queues).
• Monitoring dashboard (metrics: latency, accuracy, throughput).
• Team training (2 hours).
• 1 month support after deployment.

Model comparison for vision tasks

In our tests, Claude Sonnet 4.5 is 30% faster than GPT-4o at table parsing and gives 2% fewer errors. However, for long videos with large context, Gemini 1.5 Pro is more efficient.

To find the optimal solution for your data, contact our engineer. Get a consultation within a day. Write to us—we'll build a multimodal pipeline for your task.

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.