Building an AI Interviewer Solo: Stack, Costs, Lessons
Method: product behavior was checked against the live implementation; market and regulatory claims use the linked sources below. Dates change only after claims are re-verified.
I built Gaugely — a voice AI that ranks CVs, runs structured interviews, and drafts evidence-backed scorecards — as a solo founder, with AI coding agents writing a large share of the code. The architecture that made it possible fits in one sentence: code guarantees the interview; the model only phrases and judges.
What I built, and how big “solo” really is
Gaugely is an AI hiring team for recruiters: it ranks CVs against a rubric, runs behavioral and live-coding interviews in one voice call, and drafts a scorecard where every score quotes the candidate’s actual answer. A named human reviews the evidence and makes the decision — always.
“Solo” needs a disclosure in 2026: one person owns every decision here, but AI coding agents wrote a large share of the code. That changes what solo means. The bottleneck is no longer typing speed — it is architecture, verification, and knowing which of the agent’s confident answers to distrust. The judgment stayed the whole job; the labor compressed.
The one architecture decision that mattered
A voice demo is a weekend project now. A trustworthy interview is not — and the difference is everything a language model will not reliably do: watch the clock, ask every required question, respect consent, keep an audit trail. So the core rule of the system is a split: the orchestrator owns what must be guaranteed; the model owns only judgment.
Concretely, the model cannot skip a required question, because advancing the interview is a tool call the orchestrator validates — the model asks to move on; code decides whether it may. “Not asked” can never be scored like “answered badly,” because coverage is state the system tracks, not something the model remembers.
| Guaranteed by code | Judged by the model |
|---|---|
| The clock, and time left per stage | How to phrase the next scripted question |
| Question coverage — every required question asked | When a follow-up is worth the time |
| Consent before recording, disclosure before AI | What a candidate’s answer actually demonstrates |
| The audit trail: transcripts, scores, evidence quotes | Draft scores — each one carrying a quoted answer |
What does the stack look like — and cost?
Two runtimes, one shared database, and everything external behind a swappable port. The web app is Next.js; the interviewer is a persistent Python worker on LiveKit Agents, because live audio needs a connection that outlives a request. LiveKit carries realtime audio and control messages, Postgres carries durable state, and an authenticated completion webhook triggers web-side follow-up.
Every provider was chosen with the same test: best free tier or cheapest unit price today, replaceable tomorrow without a rewrite. That is what a port means in practice — when a model provider deprecates or a price moves, the swap is a config change, not a refactor. Marginal cost per interview lands in cents: speech-to-text lists around $0.0077 per minute (Deepgram Nova-3, late-2025 pricing), and the LLM and TTS legs add fractions more.
runtimes — a Next.js app and a persistent Python voice worker
GAUGELY ARCHITECTURE
egress fees on interview recordings
CLOUDFLARE R2
per minute of speech-to-text at list price (late-2025)
DEEPGRAM NOVA-3
EU high-risk deadline the architecture was designed for
COUNCIL OF THE EU
- →Voice loop: Deepgram speech-to-text, Groq or optional Mistral for the interviewer LLM, Deepgram text-to-speech for English or Azure Speech for supported multilingual voices, and Silero for voice-activity detection.
- →Platform: LiveKit Cloud for audio, Neon Postgres, Cloudflare R2 recordings, Azure Container Apps hosting.
- →Business: Polar as merchant of record, Resend for email, Better Auth self-hosted — the boring layer that makes it a product, not a demo.
Three production failures worth blogging
The demos never show what breaks, so here is my incident list. Each of these cost real hours and changed the design.
- →Deploys nearly killed live interviews. Azure Container Apps hard-caps graceful shutdown at ten minutes — the hour-long value in its docs turns out to be a probe field, not termination grace. A mid-interview deploy would have cut the call; the worker now drains inside that cap before a new version takes over.
- →Recording failures were silent. Managed video platforms meter recording transcode, and when a tier’s cap is hit, recordings just stop. Failures now surface on an admin dashboard instead of vanishing — the fix was observability, not budget.
- →Completion-based billing was a free-lunch bug. Interviews originally debited a team’s budget on completion — so a session that never finished never cost anything, an open invitation to drain paid capacity. Billing moved to interview start, with rate limits on top.
The biggest rewrite: my agent was a scripted stopwatch
Version one enforced structure the crude way: fixed questions, fixed timings, the model doing little more than rephrasing a script. Auditable — and robotic. Candidates would ask to clarify something and the agent would plow forward, because the script said so. Structure without conversation is just a form with a voice.
The redesign keeps every guarantee in code but hands the conversation to the model through gated tools. It can acknowledge, re-ask, follow up, or repeat a question a candidate missed — but moving to the next required question is a tool call the orchestrator validates against coverage and the clock. I am shipping this in phases, and the early lesson is blunt: the intelligence was never the hard part; the guardrails were.
Compliance shaped the product more than the demo did
AI used for recruitment is classified high-risk under the EU AI Act, with obligations applying from 2 December 2027. I treated that as a design input, not future paperwork: consent before recording, disclosure before any AI involvement, no score threshold that auto-rejects anyone, and a named human decision-maker on every hire.
The market enforces this before any regulator does. One payments provider declined us outright over the EU risk classification while I was setting up billing — the bar is real even for a tiny company. And candidates vote too: in 2026, Greenhouse found 38% of job seekers had walked away over a required AI interview, while only 19% wanted less AI in hiring. Transparent AI with a human decision is not a legal tax; it is the product.
What would I tell another solo builder?
Five things I would repeat, and one I would not.
- →Buy the real-time infrastructure. WebRTC at production quality is its own company; LiveKit-class platforms are the correct make-vs-buy call for one person.
- →Make every external service a port. Free tiers got this to market at near-zero infra cost, and no provider can hold the architecture hostage.
- →Design the failure states first. Dropped calls, half-finished interviews, evasive answers — if those states are not explicit, they become silent data loss.
- →Let AI write code, never conclusions. Agents produced huge amounts of working code and several confidently wrong designs; the verification loop is the founder’s actual job.
- →Put the compliance bar in the architecture early — retrofitting consent, audit trails, and human review into a shipped product is the expensive version.
- →The one I would skip: building the scripted version first. Structure and conversation were never actually in tension; I just assumed they were.
One founder, two runtimes, $0-egress recordings, and an EU high-risk compliance bar. The architecture, stack choices, and production failures behind Gaugely.
Inspect the scorecard it producesQuestions people ask
Can one person build an AI interviewer in 2026?
Yes, with two qualifiers: AI coding agents compress the labor dramatically, and the founder still owns architecture, verification, and every product judgment. The voice demo takes a weekend; the trustworthy version — enforced question coverage, consent, audit trails, evidence-backed scoring — is where the year goes.
What stack do you need for a production voice AI agent?
A persistent worker for the audio loop (Gaugely uses Python on LiveKit Agents), Deepgram speech-to-text, Groq or optional Mistral for the interviewer LLM, Deepgram English or Azure multilingual text-to-speech, VAD (Silero), and ordinary web infrastructure — Next.js, Postgres, object storage — around it.
How much does it cost to run an AI interview?
Marginal cost lands in cents per interview: speech-to-text lists around $0.0077 per minute (Deepgram Nova-3, late-2025 pricing) and the LLM and TTS legs add fractions more at current unit prices. The real costs are elsewhere — real-time infrastructure tiers, recording transcode limits, and the engineering time behind reliability.
How do you make an AI interviewer legally safe to use?
Build to the EU AI Act’s high-risk bar even if you are not in the EU: disclosure before any AI involvement, consent before recording, an audit trail, no automatic rejection at any score threshold, and a named human making every hiring decision. High-risk obligations for recruiting AI apply from 2 December 2027.