What changed
Three capabilities landed together: kyutai-labs/pocket-tts gives offline CPU-only speech output (FACT, source repo); Ollama shipped a first-party free web-search API for grounding local agents (FACT, source blog); and Hugging Face/Cerebras announced Gemma 4 for real-time voice AI (FACT that the announcement exists; the signal itself notes details were inferred from the title). Together they plausibly assemble an STT/LLM/TTS/live-RAG voice agent with near-zero per-minute cost (HYPOTHESIS — the full-stack latency/quality claim is not demonstrated in the sources).
Why now
Cloud voice-agent economics are dominated by per-minute STT/TTS/LLM fees; an open stack that removes them flips margin structure for always-on voice products (HYPOTHESIS derived from the signals). Counterpoint: cloud per-minute prices are falling fast and incumbents can drop prices, so the cost edge is a depreciating asset (HYPOTHESIS).
Converging signals
1) pocket-tts: CPU-only offline TTS (https://github.com/kyutai-labs/pocket-tts). 2) Ollama web search: free live-web grounding for local models (https://ollama.com/blog/web-search). 3) Cerebras+Gemma 4 real-time voice (https://huggingface.co/blog/cerebras-gemma4-voice-ai). CAVEAT (FACT from the signals themselves): Cerebras is a hosted inference service and Ollama web search is a hosted API — so the 'fully self-hosted, zero-marginal-cost' framing of this convergence is partly false; truly self-hosted real-time LLM inference needs GPU hardware the founder would have to buy and operate.
Customer pain
Missed after-hours calls cost trades (towing, HVAC, plumbing, locksmith) real jobs; human answering services are expensive and script-bound. HYPOTHESIS — the demand_evidence array is EMPTY, so no complaint, hiring, or mandate evidence supports this pain in this input.
Who pays
Owner-operators of high-callout-value local service businesses paying flat monthly for never-miss-a-call answering (HYPOTHESIS — no evidence provided). Existing spend on human answering services (Smith.ai, AnswerConnect) is known market context but was NOT supplied as evidence here, so it cannot be scored as proven.
Solved today
Human answering services ($200-500+/mo), voicemail, or cloud AI-voice platforms (Vapi, Retell, Bland, Synthflow) that charge per-minute and are usually sold to agencies/developers rather than directly to trades (FACT about competitor existence from general knowledge — flagged as uncited context, not from provided sources).
Why current solutions are bad
Human services are costly and don't scale overnight; per-minute AI platforms squeeze margins for always-on lines and require technical assembly. The self-hosted stack lets a builder sell flat-rate unlimited-minutes answering with fat margins (HYPOTHESIS).
Proposed product
A vertically-packaged, flat-rate AI phone agent for ONE trade (e.g., towing): answers 24/7, qualifies the job, quotes from the shop's rate sheet, texts the owner, books the callout. Runs on founder-controlled infrastructure (one GPU box or Cerebras while cheap) so unlimited minutes at flat price is profitable.
MVP version
Twilio/SIP number → open STT → Gemma 4 dialogue with trade-specific prompt+rate-sheet RAG → pocket-tts out; owner gets SMS summary. One pilot shop, one trade, measured on booked-jobs-captured.
30-day build
Build the pipeline end-to-end and hit sub-1.5s turn latency (the make-or-break technical gate); run it on the founder's own inbound line; recruit 3 pilot shops from one trade at $0 for 30 days in exchange for call recordings and testimonials.
60-day build
Convert pilots to $199-299/mo flat; instrument missed-call-recovered revenue per shop as the sales artifact; build the self-serve onboarding (rate sheet upload, forwarding setup).
90-day revenue plan
10-15 shops in one metro/one trade at ~$250/mo ≈ $2.5-3.7k MRR; sell with the pilots' recovered-revenue numbers. Realistic only if latency/quality gate passes and demand assumption survives contact (HYPOTHESIS).
Distribution path
Direct demo-led outreach: call shops after hours, let the AI answer their own test call back to them; trade Facebook groups and dispatch-software marketplaces later. No ad spend. Weakness: this is 1:1 founder-led sales, which the founder can do but which is slow.
Pricing hypothesis
$199-399/mo flat, unlimited minutes — the flat rate IS the differentiation vs per-minute platforms, and it is only sustainable because of the near-zero marginal cost stack.
Technical difficulty
HIGH for real-time: sub-1.5s round-trip with open STT + LLM + CPU TTS + telephony is genuinely hard engineering; quality/latency of this exact stack is undemonstrated in the sources (the Cerebras signal explicitly says details were inferred from the title). This is the primary technical risk.
Legal / regulatory risk
Call-recording consent laws (two-party states) and AI-disclosure rules for automated calls; manageable with disclosure prompts but must be built in. Low-moderate.
Platform dependency
Telephony (Twilio/SIP) is a real dependency; Cerebras pricing and Ollama's free search tier are others — the 'zero-cost' premise silently depends on two hosted services staying cheap/free, or on buying GPU hardware.
Founder fit
MODERATE at best. This is not the founder's proven wedge: no government mandate, no forced filer, no per-filing monetisation (the high-confidence portal-mandate lesson, 0.80, does not apply). It matches his automation/AI-workflow strength and demo-led selling, but real-time voice engineering and SMB churn management are new muscles. The capital-and-runway lesson (0.9) means the GPU box and 3-6 month ramp are NOT disqualifying.
Breakout potential
If one trade works, the playbook replicates across trades and into kiosk/coaching verticals; but the cost edge erodes as cloud prices fall, so breakout depends on vertical depth, not the stack.
Final recommendation
PASS for now — do not build. The convergence is real as a capability shift, but this input contains no demand evidence, the founder-fit is mediocre (no forced buyer, no portal mandate, crowded market, cost-based moat), and the core 'zero marginal cost / fully self-hosted' premise is partly false. Revisit only if the engine later surfaces PAIN or HIRING evidence for missed-call answering in a specific trade; then run a 2-week demand probe before any build.
Next action
Queue a demand-side scan: search complaints/job postings for 'missed after-hours calls' / 'answering service' pain in 2-3 candidate trades (towing, HVAC, locksmith); only re-score this opportunity if concrete demand evidence attaches.