What I Built AI voice agents designed with a practicing PI attorney. Every question reflects how an experienced lawyer evaluates the strength of a case. Dynamic case intake. The agent identifies new cases, existing clients, or other inquiries and routes accordingly. For new cases, it asks the right questions adapted to case type — auto accidents get different questions than medical malpractice. Bilingual conversations. The system detects language early and conducts the entire intake in English or Spanish — tested across Mexican, Caribbean, and Central American dialects. Document collection. During the call, the system sends an SMS with a secure upload link. Vision AI analyzes documents and adjusts case scoring in real-time. Case scoring. Based on conversation and documents, the system assesses case strength — fault clarity, injury documentation, and insurance situation. Scoring learns from attorney feedback over time. Intelligent transfers. When criteria are met, the system transfers live to the attorney with full context briefing. Turn-based voice architecture. We chose speech-to-text, then LLM processing, then text-to-speech over real-time models. This produces clear audit trails for legal compliance, scales cost-efficiently, and allows swapping components as better models emerge. Multi-provider resilience. An orchestration layer switches LLM, TTS, and STT providers per company, per agent, per call type at runtime. Automatic fallbacks keep calls stable when services fail. Per-firm customization. Each firm's intake flow differs significantly. The system uses custom question banks at global, company, and agent levels with dynamic prompt injection. Quality evaluation. LLM-based framework measures response quality, cost, and latency on every call. Regressions surface immediately.