Radmard

Company AI workspace: one login for chat, agents, and document tools. SSO, usage budgets, and DLP.

• Sized for about 130 people online at once on live chat. Stress-tested past that without a crash.
• Shipped as a full product: web + Android, Docker/CI on Linux.

Persian + Arabic religious RAG over a library of 50+ books, each 800+ pages. Answers come from the books, not a naked LLM.

• Owned the full ingest: scan/OCR → clean text → chunk → vector indexes.
• LlamaIndex + LlamaCloud: hybrid dense + sparse, multi-query fusion, RRF, then one global Cohere rerank. Same pattern for law and Sepidar corpora.

A team records a Farsi meeting. They leave with minutes, actions, and a follow-up email, not a raw WAV.

• Upload, microphone, or tab capture (Meet / Zoom / Teams). ElevenLabs Scribe captions plus a speaker-labeled transcript.
• An LLM writes the pack: summary, decisions, action items, follow-up mail. Keycloak login.

Speech-to-speech agents with RAG and in-call tools. Latency, barge-in, and call quality are the job.

• EngageIQ at EuroShop: QR → talk → tools capture the lead. LiveKit / Vapi realtime, ElevenLabs TTS.
• DFKI lab avatar: every answer hits vector RAG first. Beauty Lounge / Morphoderm: barge-in, language switch, conversation QA.

Finance drops a scan. They get fields they can approve, or the file never leaves the building.

• PP-OCRv5 on GPU when data stays in-box; cloud VLM when it can leave. On-prem extract uses a quantized local LLM (Ollama / Qwen).
• Human-in-the-loop review. Invoice schema (parties, line items, bboxes). Batch sefte → Excel.

Staff type a problem in Farsi. FAQ answers first; if that fails, an LLM triages until a human takes over.

• Reads screenshots, sets priority, escalates on policy. Manager dashboard for queue load and handle time.
• Fast ticket create: the AI runs in the background so the user is not staring at a spinner.