Back to profile

2025 — 2026

Speech & AI workflow platform

Meeting recording, transcription and AI processing in one system. From conference bots to custom LLM workflows and local inference.

Architect and primary developer of AI subsystems

The challenge

Separate processing modules duplicated code and competed for compute resources. Each new AI feature required its own implementation.

What I built

  • Built recording bots for Zoom, Google Meet and Yandex Telemost, with a Docker container for each meeting.
  • Unified transcript processing and research jobs in a shared LLM Task engine.
  • Developed multi-step workflows, structured responses and long-recording processing.
  • Built a local-model gateway, telemetry and management interfaces; integrated transcript chats.

Results

  • Three meeting platforms connected to one recording and processing workflow. (3 platforms)
  • Different AI tasks run through a shared engine and resource pool.
  • Tasks, models and servers can be monitored through management interfaces.
System architecture. Recording, transcription, AI tasks and local models form one flow. Solid lines show the data flow; dashed lines show feedback and operational connections.Open full-size diagram ↗
Components & connections

Sources

  • Zoom · Meet · Yandex TelemostMeeting sources.

Recording & transcription

  • Recording botsOne isolated container per active meeting.
  • TranscriptionText and timing from a recording.
  • Recording queueAsynchronous media processing.
  • Audio extractionPrepare audio for transcription.

LLM orchestration

  • Universal LLM TaskShared execution model for AI tasks.
  • Parent / child tasksSplit long context and aggregate results.
  • Scenario engineSequential and conditional AI workflows.
  • Structured outputParse model responses into typed results.
  • Prompts & assistantsInstructions and response configurations.

Local inference

  • LLM gatewayEntry point for local inference and telemetry.
  • llama.cppRuntime for local open-weight models.
  • Load balancerRoute requests to available inference endpoints.
  • Open-weight modelsLocally hosted language models.

Product interfaces

  • Operations dashboardTasks, reports, recordings and infrastructure status.
  • Transcript chatsQuestions about individual recordings.
  • Local agentContainerized agent using local inference.
  • External CRMRecording, transcript and structured report integration.

Operations

  • TelemetryInference, task and server status.
  • SSH orchestratorInstallation, updates and health workflows.
  1. Zoom · Meet · Yandex Telemost→Recording botsmeeting link
  2. Recording bots→Recording queuerecording
  3. Recording queue→Audio extractionmedia
  4. Audio extraction→Transcriptionaudio
  5. Transcription→Universal LLM Tasktext context
  6. Universal LLM Task→Parent / child taskslong context
  7. Prompts & assistants→Universal LLM Taskdefinition
  8. Scenario engine→Universal LLM Taskspawns
  9. Universal LLM Task→Load balancerinference request
  10. Load balancer→LLM gatewayroutes
  11. LLM gateway→llama.cppproxy
  12. llama.cpp→Open-weight modelsruns
  13. Universal LLM Task→Structured outputresult
  14. Structured output→Operations dashboardreport
  15. Structured output→External CRMtyped data
  16. Transcription→Transcript chatsconversation context
  17. Transcript chats↔Local agentprivate mode
  18. Local agent↔LLM gatewaylocal inference
  19. LLM gateway→Telemetrysignals
  20. Telemetry→Operations dashboardstatus
  21. SSH orchestrator→Recording botsdeploy/update
  22. SSH orchestrator→LLM gatewaydeploy/update

Engineering takeaway

Context

A platform for meeting recordings and local language models.

  • Data processing within a local environment.
  • Varying availability of compute nodes.
  • Recordings longer than a model’s context window.

Approach

Extracted a shared engine from existing features, then progressively connected recording, processing, workflows and infrastructure.

  • One task model and a shared resource pool for different AI features.
  • Child tasks and aggregation for long context.
  • Separate models for prompts, workflows, execution and reports.

Lessons

  • Generalization is justified once duplication has been demonstrated, rather than in advance.
  • Structured responses connect model output to application contracts.
  • Observability and deployment are part of the product, rather than ancillary work.

Stack

  • Docker
  • Queues / workers
  • llama.cpp
  • Open-weight LLM
  • Nginx / load balancing
  • SSH orchestration
  • Vue dashboards
  • Structured output

Skills

Products, systems and ideas brought to life in working software.