CallingBox vs Gemma 4 Multimodal Fine-Tuner

CallingBox is a YC-backed API that makes AI phone calls a one-liner. You configure a reusable agent with instructions, persona, and tools — then dispatch outbound or inbound calls via a single endpoint. The AI conducts the full conversation, then returns structured JSON matching whatever schema you defined. No managing telephony stacks, STT, TTS, or LLM pipelines separately. At $0.05 per connected minute all-inclusive — covering telephony, speech-to-text, language model, text-to-speech, and data extraction — it's substantially cheaper than stitching together LiveKit, Deepgram, GPT-4o, and ElevenLabs yourself (which their own benchmarks put at ~3x the cost). Sub-500ms latency with a 4.31 MOS quality score makes it production-ready. IVR navigation, voicemail detection, DTMF support, and MCP server integration cover the tricky edge cases that kill most voice implementations. Founded by Jonathan Chávez and Sebastian Crossa, the company offers $5 in free credits to get started. The use cases are obvious and immediate: appointment reminders, collections, customer support, multilingual outreach. For any team that's been putting off voice because of infrastructure complexity, CallingBox removes the excuse.

Gemma 4 Multimodal Fine-Tuner is an open-source toolkit that lets developers fine-tune Google's Gemma 4 and 3n models across all three modalities — text, images, and audio — using only Apple Silicon hardware. It runs natively on PyTorch with Metal Performance Shaders (MPS), bypassing the NVIDIA requirement that has historically blocked Mac users from serious local fine-tuning work. The toolkit handles the full training pipeline including dataset prep, LoRA adapters, and multi-modal data collation. It ships with working example notebooks, a validation suite, and clean abstractions that don't require deep familiarity with the underlying MPS stack. Apple Silicon's unified memory architecture actually helps here — large multimodal batches fit in memory that would otherwise require GPU VRAM splitting on CUDA setups. Posted to Hacker News on April 7 as a Show HN, it pulled 109 upvotes and 165 GitHub stars within hours. The timing is sharp: Gemma 4 just dropped days ago with new multimodal capabilities, and the community immediately wanted local fine-tuning. This fills that gap faster than Google's own tooling.