How do you architect one AI platform that runs identically in the cloud and inside a private research lab?
Azati architected and built the AI platform powering a biotechnology and intellectual property analytics company's patent and sequence research tools. The engagement centered on system architecture: a hybrid-ready microservices design with open-source cloud parity, a pluggable LLM layer, AI pipeline orchestration, and hybrid vector search, built to run identically on cloud infrastructure or fully on-premise for research environments with strict data governance needs.
reduction in manual annotation time
semantic search accuracy across patent and sequence data
hybrid-ready deployment across cloud and on-premise infrastructure
A biotechnology and IP analytics company needed to turn a flood of unstructured patent documents and biological sequence data, spread across PDF, XML, FASTA, and GenBank formats, into something researchers and IP analysts could actually search, understand, and act on quickly. The functional need was clear. The harder problem was architectural: the platform had to work for research teams operating on public cloud infrastructure and for clients whose data governance requirements meant the system had to run entirely on-premise, with no cloud dependency at all, using the same codebase.
Azati architected the platform around that constraint from the start: a microservices design built on open-source components with direct cloud-service parity, a pluggable LLM layer that supports open-weight models, and orchestrated AI pipelines for ingestion, enrichment, and search, all deployable identically through Docker and Kubernetes regardless of where the infrastructure actually lives.
Patent documents and biological sequences arrived in fundamentally different formats, PDF and XML for patents, FASTA and GenBank for sequences, updated continuously and requiring standardized ingestion:
Simple keyword matching could not support the kind of research queries patent analysts and scientists actually needed to run:
Patent claims and sequence annotations are dense, technical, and often reference other documents, which made naive summarization unreliable:
The platform had to be deployable in environments the client did not fully control and could not assume would ever have cloud connectivity:
The platform is decomposed into independently deployable services rather than a single monolithic application, letting ingestion, search, summarization, and orchestration scale and evolve separately without one component's load or failure affecting the others.
Every cloud-native capability in the platform, object storage, relational data, search and vector indexing, has a self-hosted open-source equivalent in active use: MinIO in place of proprietary object storage, PostgreSQL for structured data, Elasticsearch paired with a vector database for search. The same platform runs on public cloud infrastructure or entirely within a client's own data center.
Luigi coordinates the ingestion, enrichment, embedding, and indexing pipelines as managed, observable workflows rather than as a collection of scripts run manually or on ad hoc schedules.
Elasticsearch handles traditional full-text search while a vector database handles semantic similarity over LLM-generated embeddings, combined into a single hybrid search layer that matches both exact terminology and conceptual meaning.
Docker and Kubernetes containerization, provisioned through Ansible, make the deployment target, cloud, on-premise, or a mix of both, a configuration choice rather than an architectural rewrite.
A model abstraction layer sits between the application logic and the underlying language models, currently LLaMA and Mistral, so models can be swapped or fine-tuned for a specific research domain without touching the rest of the system.
Deployment-agnostic architecture, open-source cloud parity, and pluggable LLMs are what let this platform run identically anywhere. Let's talk about your infrastructure constraints.
Architect my AI research platformFive modules cover the platform end to end, from conversational research assistance down to the storage and infrastructure layer everything else runs on.
A conversational research assistant built on a multi-agent pattern using LangChain, letting researchers query patent and sequence data in natural language and receive answers grounded in the platform's underlying search and retrieval layers rather than a generic language model response.
Automated summarization of patent claims and sequence annotations using template-driven prompting and fine-tuned LLMs, producing consistent, structured summaries instead of free-form text that varies in format from one document to the next.
Machine learning-based classification and clustering applied across the ingested dataset, along with cross-domain entity recognition that links related patent claims and sequence records that would otherwise remain disconnected.
An orchestration framework connecting the platform's internal services, using RabbitMQ for asynchronous task distribution and Keycloak for centralized authentication and authorization across every service in the platform.
The storage and infrastructure foundation the rest of the platform depends on: PostgreSQL, MinIO, and Elasticsearch for data persistence and search, deployed and scaled through Kubernetes across whatever environment the platform is running in.
| Area | Azati contribution |
|---|---|
| System architecture | Designed a microservices-based, deployment-agnostic platform architecture |
| Cloud parity | Built open-source equivalents for every cloud-native service used |
| LLM layer | Implemented a pluggable model abstraction layer supporting LLaMA and Mistral |
| Pipeline orchestration | Built Luigi-based orchestration for ingestion, enrichment, and indexing |
| Search | Implemented hybrid full-text and vector search across patent and sequence data |
| Deployment | Containerized the platform for identical cloud or on-premise deployment |
| Access control | Implemented Keycloak-based authentication and authorization |
Access control across every service in the platform runs through Keycloak with OAuth2, giving the client centralized authentication and authorization rather than per-service credentials. For clients with strict data governance requirements, the platform's open-source cloud parity means the entire system, data storage, search, and AI inference, can run fully on-premise with no data ever leaving the client's own infrastructure.
Automated metadata enrichment and classification took over work that previously required manual review of patent claims and sequence records one at a time.
Hybrid full-text and vector search across patent and sequence data consistently returns results that match on meaning, not just matching terminology, which is what a 90%+ accuracy figure actually reflects in a domain this technical.
The same platform, unmodified, runs on public cloud infrastructure or entirely on-premise, giving the client flexibility to serve research clients with different data governance requirements without maintaining two separate systems.
The model abstraction layer means the platform isn't locked to LLaMA and Mistral specifically. As better or more specialized open-weight models become available, they can be integrated without rearchitecting the application layer around them.
What this engagement demonstrates:
A platform built cloud-first and later ported to on-premise almost always carries architectural assumptions that don't survive the transition. Designing for open-source cloud parity from the first architectural decision is what made identical cloud and on-premise deployment achievable rather than aspirational.
Full-text search alone misses conceptually related results phrased differently. Vector search alone can drift from exact technical terminology that matters in patent and sequence work. Combining both into one hybrid layer is what got semantic search accuracy above 90% in a domain where precision actually matters.
Open-weight models capable enough for domain-specific fine-tuning are improving quickly. Building the application layer against a model abstraction rather than a specific LLM is what keeps the platform from needing a rearchitecture every time a better model becomes available.
Explore our successful projects and see how Azati delivers measurable results for our clients.
The client needed to process massive volumes of unstructured patent documents and biological sequence data, spread across PDF, XML, FASTA, and GenBank formats and updated continuously, while ensuring high-quality metadata, scalable processing, efficient retrieval, and compliance with global IP standards. Processing terabyte-to-petabyte-scale datasets required large-scale ingestion, normalization, cleaning, and indexing pipelines capable of handling diverse and unstructured data, automating workflows that were previously labor-intensive and manual.
Azati developed an AI-driven platform that standardizes heterogeneous datasets, removes duplicates, cleans corrupted records, and prepares them for AI-powered search and metadata enrichment. A dedicated enrichment module restores lost connections between sequences, annotations, and patent claims, while a semantic and similarity-based search layer built on Elasticsearch, vector databases, and LLM embeddings enables fast, context-aware retrieval across patents and sequences.
Pharmaceutical marketing teams needed to correlate physician questionnaire responses and field feedback with marketing strategy performance, a process that was manual, slow, and difficult to scale across large volumes of unstructured text and reporting materials. Generating consistent, evaluable marketing assessment reports from this data required both language understanding and visual document processing working together.
Azati developed an AI and computer vision solution that automates the analysis of physician questionnaires and correlates feedback patterns with marketing strategy outcomes. The system combines NLP-based text mining with automated report generation, producing comprehensive evaluation reports for pharmaceutical marketing teams without manual data correlation.
A biotechnology corporation providing access to genomic sequence information struggled to get meaningful results when running short DNA/RNA primer sequences, under 20 bases, through the standard BLAST algorithm. The default configuration prioritized longer sequences, causing missing matches and incomplete alignments that slowed genomic analysis and undermined primer-based research critical to personalized medicine work. Short sequences also carried a heavy risk of false negatives due to BLAST's probability-driven scoring model, leaving biologically meaningful partial matches undetected.
Azati's engineering team implemented changes to the BLAST algorithm's codebase to support dynamic parameter tuning based on input sequence length, along with enhanced filtering and sensitivity tuning to surface previously hidden alignments without cluttering results with noise. The customized configuration lowered significance thresholds specifically for short sequences and automated the tuning process so researchers no longer needed to manually configure parameters for each primer-based search.
A bioinformatics company needed a search engine capable of understanding the actual meaning behind research queries across vast, complex scientific datasets, rather than matching on exact keywords. Multiple naming conventions and variations in disease names across the data hindered precise tagging and search accuracy, and the project had no pre-labeled sample data available for algorithm training, a significant obstacle for building an accurate matching system from scratch.
Azati built the training dataset from open-source life sciences documents, analyzing hundreds of thousands of records to construct a comprehensive thesaurus and training a custom Word2Vec model to detect and map synonyms across disease naming conventions. The team developed intelligent matching and query analysis modules and implemented RESTful microservices with Redis caching for scalable, efficient search performance, along with a pluggable module for automatic tagging of blood samples based on their descriptions.
Last updated
