Phonely

Telephony-first latency engineering: multiple contradictory latency claims (~1s latency for voice API vs sub-400ms LLM responses) imply a tiered, mixed pipeline—fast lightweight models or heuristics for immediate replies plus heavier fine-tuned models for richer responses. That design (fast-path controller + slow-path full model) is a pragmatic pattern for voice where perceived latency matters.

Deep SIP/carrier-level engineering: explicit mention of SIP REFER, post-transfer minute accounting, and enterprise custom telephony signals they’ve implemented a fairly complete SIP stack and carrier interconnect logic. Handling transfer semantics to avoid carrier charges is a subtle, high-friction engineering area most AI-first startups ignore.

Unified conversation flow model across channels (voice, SMS, webchat) with the same flow semantics and business logic. Rather than separate bots per channel, they reuse call flows and routing rules, which requires a channel-agnostic state machine, message normalization, and media-aware adapters (ASR/TTS for voice, text endpoints for SMS/webchat).

Industry-specific fine-tuned LLMs + BYO model support: offering 15+ fine-tuned models for verticals and 'support for all major LLMs' suggests an orchestration layer that can host, switch, or ensemble models (likely quantized or distilled) depending on compliance, latency, or domain.

Real-time streaming API + webhooks + call path visualization + A/B testing: demonstrates non-trivial instrumentation — streaming transcripts, evented state, and per-branch analytics — so they’ve built traceable, experimentable call state that can be replayed/visualized for optimization and regulatory audits.