Phonely's Custom LLMs Fool 80% of Callers on Millions of Calls
Phonely handles millions of calls/month across hundreds of verticals using modular custom LLMs that optimize outcomes statistically—e.g., one question tweak boosts results 5%—fooling 80% of callers into thinking it's human.
Modular Custom LLMs Beat Off-the-Shelf Models for Voice AI
Phonely's core edge comes from building small, specialized LLMs instead of relying on general models like OpenAI's. Break voice AI into components—e.g., one model stores variables like customer names or emails, another handles other tasks—running on fast inference hardware like Groq to cut latency while matching quality. This modular setup isolates updates, saves costs, and maintains low latency (now "good enough" like oxygen, shifting focus to conversational quality and accuracy). Result: AI handles edge cases like garbled audio, interruptions, and endpointing, performing well enough that 80% of callers have no idea it's not human—rising to near 100% by year-end.
Optimization drives outcomes over mere conversation: Surface call data to customers, revealing tweaks like changing one question to boost performance 5%. With millions of calls/month across hundreds of verticals (call centers, insurance, home services), this data moat statistically improves agents on revenue metrics, like qualifying leads from billboards or booking appointments without human handoff (unless regulated, e.g., licensed insurance agents).
Differentiation from voice AI boom: Early PhD experiments yielded battle-tested production knowledge, plus telephony expertise generic models lack. Prioritize inbound revenue calls (sifting high-value leads) over support; disclose AI for outbound due to ethics/regulations, but inbound users prefer context-aware AI over random humans.
Pivot to Enterprise Call Centers for Explosive Revenue
Start with small businesses ($30-100/month) for fast feedback and iteration (4-5 months), then pivot when one call center outpays all SMBs combined. Target high-volume users caring about lead qualification and optimization—e.g., home services or insurance running call centers. This unlocked scale: millions of calls/month, $16M Series A led by Base10 Partners (sparked by founder's LinkedIn post on ultra-endurance cycling lessons in commitment).
Future: 50M+ calls/month, hiring sales/growth and low-ego engineers in SF. Voice AI moat deepens via proprietary data and telephony, outpacing generic models.
Founder Reality: Endless Daily Battles, Create Your Own Luck
From athlete (cross-country skiing, ultra-cycling: 300+ mile non-stop races) to failed audio startup to AI PhD in Australia, inspiration hit watching dad's practice struggle with phones. Key lesson: Success looks overnight on LinkedIn/Twitter but feels like constant war—new models, competitors, daily proof. Set expectations: Fight every day.
Advice: If you "want" to found, test at a startup first. If no choice, roll the dice relentlessly—create luck through volume. Pick investors like hires (e.g., Base10's excitement led to preemptive Series A).