Mehran Toosi · Principal Software Engineer · Copenhagen

Blog

How systems I built work, one step at a time. Each step has an animation of that part of the system.

· 13 min readOrvena 4B: training a 4B model to do phone tasks without thinkingHow I fine-tuned Qwen3.5-4B to do Orvena’s phone tasks with thinking off, with a bench that runs the app’s own agent loop, training data filtered by its checks, and the models that failed on the way. · 8 min readCue: a small on-device model for turn-taking in voice modeHow Cue-0.4B decides, in 30 languages and about 10 ms on the iPhone Neural Engine, whether speech during an answer means keep going, stop or a new request. · 5 min readRunning an agent with a 4B model on an iPhoneHow Orvena fits a 4B model and 38 tools into an iPhone’s memory, keeps its prompt cache, repairs a small model’s tool calls and tests every change on the phone. · 5 min readRouting an agent with four small modelsHow ecoai picks the model, tools and first step for each message with four small TF-IDF and logistic-regression models that stay silent when unsure. · 6 min readKeeping customer data from the model without blinding the agentHow Privacy Guard swaps personal data for stable tokens before a model call and restores them before a tool runs, so an agent acts on data its model never sees. · 4 min readA knowledge graph that lets an agent answer data questions correctlyHow ecoai’s knowledge graph is built from the product schema, ETL pipelines and five layers of SQL, then searched, prepared and checked before SQL runs. · 3 min readCW Luma 10: a denser visual code that stays easy to scanHow CW Luma 10 (ChromaWeave) puts ten bits in each cell where phone cameras see detail best, checks its own reads, and moves files from a screen to a phone.