Case studies
Dialogue systems on top of large language models: answers drawn strictly from the client’s knowledge base, deterministic scenarios instead of free generation, and hand-off to a human for hard cases.
6 projects in this category · ML, LLM, RAG and analytics
Task: Remove routine first-line consultation from the lawyer without the risk of legal hallucination.
Solution: A deterministic state machine on top of the LLM: an 8-field questionnaire drives the dialogue order, legal answers are generated only from the knowledge base index, and 14 stop factors escalate the client to a human.
Result: Metrics: acceptance suite of 110 cases, PASS 100% against a 95% acceptance threshold, no critical violations; ~980 automated tests; a legal answer without a source citation is technically impossible.
/ai-lawyer-hero.webp)
LLM · RAG
Context: A conversational agent for a law firm: it qualifies clients along the sales script, answers legal questions strictly with a citation to the source, and hands complex cases to a human lawyer. Built from scratch in Python.
Task: Give people one window for hundreds of AI models and ready-made agent assistants, available from the browser and a phone without installing anything on every computer, and turn it into a service with clear pricing.
Solution: The logic moved to the server: every account has its own chats, folders and settings, and sign-in is protected by two-factor authentication. The interface was adapted for phones and home-screen installation. Users build agents for their own tasks, with instructions, knowledge files and a set of tools (search, documents, calculations, external services). Models search the web, and the “Research” mode plans its own questions, reads the pages it finds and assembles a final report with links to the sources. Notebooks with podcasts, memory and scheduling were added. Payment: a balance with online top-up charged by actual usage, or an “Unlimited” subscription.
Result: The desktop app became a cloud service. Hundreds of models for text, images, audio and video work in one window from a phone and a computer, agents take over routine work, and payment is transparent: the user can see what every charge was for.
/synedria-hero.webp)
Web · SaaS · LLM
Context: A cloud service that grew out of a desktop app: hundreds of models for text, images, audio and video in one window, in the browser and on a phone, installable to the home screen without an app store. Online balance top-up, charging by actual usage, an “Unlimited” subscription and a separate admin console. It offers model comparison, agents, research across sources, notebooks with podcasts, memory and scheduled tasks.
Open project ↗Task: Speed up access to team knowledge.
Solution: RAG assistants with quality control.
Result: Metrics: median response time 18s -> 6.4s (x2.8), inaccurate answers down ~43%.
/%D0%A1%D0%BD%D0%B8%D0%BC%D0%BE%D0%BA%20%D1%8D%D0%BA%D1%80%D0%B0%D0%BD%D0%B0%202026-03-07%20203310.webp)
LLM · RAG
Context: LLM assistant with RAG: drives the dialogue script and answers from a knowledge base with business context.
/%D0%A1%D0%BD%D0%B8%D0%BC%D0%BE%D0%BA%20%D1%8D%D0%BA%D1%80%D0%B0%D0%BD%D0%B0%202026-03-08%20104646.webp)
LLM · RAG
Context: Enterprise search and assistants on client data.
Task: Remove routine work from sales and support.
Solution: AI dialogue agents and request orchestration.
Result: Metrics: ~68% of routine requests closed automatically, operator load down ~37%, first response about 25 sec.
Agents and bots
Context: Orchestration and dialogue scenarios.
Agents and bots
Context: Process and communication automation.
Describe the task, the data and the timeline. We will propose a format and an outline plan. Reply within one business day, NDA on request.
Discuss a project