Case studies

Software engineering case studies, measured in uptime, latency and load.

Selected work in healthcare, security, HRTech, retail and AI: the problem, what we changed and the measurable result.

AI · LLM cost & observability

Cutting LLM cost by 60% in production while keeping quality, with full observability.

We run LLM workloads through OpenRouter and track regular and reasoning tokens, model response time and whether answers match the expected schema. Cost came down through cheaper models where they are good enough, caching and prompt corrections, and quality is checked by a second model plus test benches that run the same data through different models many times. Queues and a fallback model keep the service up, and sensitive data is masked or replaced with placeholders before it reaches a model. The hard part was balancing price against quality.

OpenRouter
Token tracking
Model evaluation
Fallback models
Data masking

Ask about this case →

−60%

LLM cost with quality maintained

AI · Knowledge assistant (RAG)

An internal assistant that answers questions from 50,000 pages of technical manuals.

We built an internal knowledge assistant over about 50,000 pages of PDF installation, operation and service manuals for technical equipment, using vector search over a vector database and the client's own model. Every answer is verified by an LLM, and quality is measured against a test set of questions and answers. Search quality and document chunking were the hardest parts. The client got an assistant with a ready knowledge base of its equipment documentation.

RAG
Vector search
Vector database
LLM verification
Evaluation set

Ask about this case →

50,000

pages of PDF documentation searchable

AI · LLM process automation

LLM-powered automation of reports, email triage, documents and support on one shared infrastructure.

We built LLM-driven components that take over routine work: preparing reports, triaging incoming email, handling documents and answering support requests. They connect to databases, mail, a CRM and internal services, with different cloud models chosen for different tasks. They are deliberately constrained: they read data and issue standard answers based on specific templates, and they never change records. The hard part was building a single infrastructure for all of them, and the client got high-quality automation of its processes.

LLM
Process automation
Multi-model
Prompt engineering
CRM & email integration

Ask about this case →

Healthcare · AI document processing

Medical documents from scans, PDFs and free text, standardized with an LLM.

Lab results, discharge summaries and doctors' orders arrived as scans, PDFs and free text, in a different format and language from every laboratory. A mid-size LLM, chosen for speed and quality per cost, ran inside the client's AWS account and normalized them to LOINC at ~5,000 documents a month. A second, stronger model cross-checked every result, and any mismatch went to manual review and stayed out of the data until resolved. The structured data now powers patient analytics, test-history comparison and trend detection.

LLM
LOINC
AWS
LLM evaluation
Data normalization

Ask about this case →

>95%

extraction accuracy on our evaluation bench

HRTech · Product build

A hiring module with AI, built from zero inside a multi-tenant HR platform.

We built the Hiring module of a multi-tenant HR platform from scratch: vacancies, candidate cards, the hiring funnel and statuses, evaluation, communication and reporting. On top of it we added resume parsing, an AI candidate analyzer and AI ranking, calendar integration, and automated emails through mail services. Anyone with the right permissions in any company on the platform can use it, so the client got hiring and the rest of HR in one all-in-one platform, with the routine work automated.

Multi-tenant
AI ranking
Resume parsing
Calendar integration
Email automation

Ask about this case →

Facing a similar problem? Let's talk specifics.

Book a Technical Consultation →