Architecting complex software systems and advanced AI.
Senior AI Technical Lead / CTO
Selected work
Booq — Persian-first AI video translation and dubbing
Booq is a live, Persian-first AI platform that brings the work of translators, voice actors, studios and editors into one web app. As CTO, Mohammad Panahi defined its architecture and took it from first commit to production: video translation and dubbing, audio, music, images and film, paid from one wallet.
Kartez Contract — AI contract drafting, review and editing
Halted by the client one month from completion, for war-related funding reasons. The engineering write-up below is the part that outlived the contract.
Sino — industrial AI, from zero to customer sale
Three years taking industrial AI products from zero to a delivered, sold product, and optimising the models to fit hard industrial and cost constraints. The work spanned large language models, graph neural networks and computer vision.
Persian text-to-speech research for MCI
Persian text-to-speech research, commissioned and supervised by MCI (Hamrah-e Aval) under Iran's National Elites Foundation. The hard part is linguistic: Persian script leaves short vowels and the ezāfe unwritten, so pronunciation has to be decided from context.
Medical imaging — MRI denoising and gene expression from pathology
Speckle-noise reduction in MRI for a private company, and an MSc thesis at Isfahan University of Technology on recovering gene-expression signal from digital pathology images, combining graph neural networks, vision foundation models and adapted language-model techniques.
Barzaman — an offline-first planner and calendar across three calendars
Barzaman is a live, offline-first personal planner and calendar for students and Konkur candidates, teachers, medical staff, engineers and managers. Mohammad Panahi founded it and leads its product and engineering. It runs natively on Android and Windows and on the web, keeps the user's data on the device, and syncs between devices only when the user chooses an account.
Writing
Hybrid search for Persian RAG: a field guide
Hybrid search for Persian (Farsi) RAG on PostgreSQL and pgvector: safe normalisation, RRF over dense and lexical ranks, no locale filter, bilingual tests.
RAG hallucination guardrails: a five-layer design
Five guardrails that stop a RAG assistant inventing answers: refusal before retrieval, a similarity floor, the prompt, a sentence gate and voice rules.
Choosing a pathology foundation model in 2026
No single pathology foundation model wins. What benchmarks, licences, task fit and compute say about UNI, CONCH and Virchow2. Research use, not diagnosis.
Streaming LLM guardrails: you can't unsay a token
Streaming LLM guardrails: a token can't be unsent. How NeMo, Guardrails AI and Bedrock trade speed for safety, and how a sentence gate checks each claim.
Legal AI hallucination cases pass 2,000: what to build
Court decisions on hallucinated AI content have passed 2,000. Why RAG alone did not fix it, and design patterns from my work on contract AI.
Persian token tax: what non-English LLM apps cost
Persian needed 1.53× the English tokens on GPT-4o's tokenizer and 3.26× on GPT-4's, across 1,012 parallel sentences. Method, results and what to test.