Stack Sensei — AI Voice Mentor Platform
Voice-Native • Emotion-Aware • Failover-Resilient
Real interviews are spoken, not typed. Stack Sensei is a voice-native AI mentor that interviews you out loud: a three-tier TTS failover chain, emotion-aware speech synthesis, and 13 distinct mentor voices deliver adaptive technical questioning across TypeScript, Java, Linux, and security. An eight-provider bring-your-own-key LLM layer keeps the intelligence flexible, while exam mode grades your answers and issues one-year certificates.
Interview Prep Is Silent. Voice AI Is Fragile.
Technical interviews are spoken conversations under pressure — yet most prep tools make you type, and most voice stacks collapse the moment a single vendor hiccups.
Text-Only Practice, Spoken Interviews
You rehearse by typing, then get judged on how you explain architecture out loud. The core interview skill never gets trained.
Single-Vendor Fragility
Most voice apps hardwire one TTS API. One outage or rate limit and the session goes silent mid-answer — exactly when trust matters most.
Flat, Robotic Delivery
One-size-fits-all synthesis sounds the same whether you nailed the question or bombed it. No encouragement, no energy, no mentor.
LLM Lock-In
Hardwiring a single model means one vendor's pricing, one vendor's latency, one vendor's outages — with no user control over keys or cost.
A Voice-Native AI Sensei
Stack Sensei 🥋 treats voice as core infrastructure, not a bolt-on: a health-monitored failover chain, emotion-mapped synthesis, and a provider-agnostic intelligence layer keep the conversation flowing.
Converse, Don't Type
A full voice loop: browser capture with voice-activity detection, ElevenLabs Scribe transcription with Whisper fallback, and a sensei that asks adaptive questions and speaks the follow-up back.
Three-Tier TTS Failover
ElevenLabs WebSocket streaming as primary, self-hosted Kokoro on Fly.io's private network as secondary, OpenAI TTS as tertiary — with active health monitoring and metrics on every hop.
Emotion-Aware Synthesis
An EmotionContext engine maps five states — from supportive to extremely emotional — to distinct synthesis settings. The sensei sounds encouraging when you struggle and fired up when you deliver.
Bring Your Own Model
Eight LLM providers behind one abstraction — Claude, GPT-4o, Gemini, Mistral, DeepSeek, Qwen, Kimi, or local Ollama — with per-user keys, key validation, and per-user circuit breakers.
Voice as a First-Class System Layer
Four layers, one goal: spoken answers in, spoken mentorship out, with no single point of failure between them. Kotlin and Spring WebFlux keep the pipeline reactive end to end.
Client Layer
React 18 single-page app where learners speak or type. Voice-activity detection knows when you've finished answering; streamed audio starts playing before synthesis completes.
Voice Pipeline
The signature layer: a three-tier TTS failover chain with health monitoring, a text preprocessor that makes technical content speakable, and dual-provider STT. Thirteen mentor voices each carry a Kokoro fallback ID so the persona survives failover.
Intelligence Layer
An eight-provider BYOK abstraction (Claude Sonnet 4.5 by default, down to local llama3:8b) drives adaptive questioning, answer grading, and conversation branching, grounded by a 61-document curated knowledge base with keyword-relevance retrieval.
Platform Foundation
Kotlin on Spring Boot 3.2 WebFlux with OAuth2 security and Caffeine caching; PostgreSQL 16 evolved through 10 Flyway migrations. Stripe handles payments, Sentry watches production, Fly.io runs it all — and jsPDF renders the certificates.
The Voice Round-Trip
Every exchange travels a four-stage loop. Each stage has a fallback, so the conversation never dies mid-sentence.
Listen
Browser → STTWhat Happens
- Voice-activity detection ends the take automatically
- Audio transcribed by ElevenLabs Scribe
- OpenAI whisper-1 stands by as fallback
- Text input always available as a first-class path
Think
LLM LayerWhat Happens
- Your chosen provider grades the answer
- Adaptive engine picks the next question by difficulty
- Conversation branching explores follow-up paths
- 61-doc knowledge base grounds the domain
Speak
TTS Failover ChainWhat Happens
- Text preprocessor makes code and markup speakable
- EmotionContext tunes stability, similarity, style
- ElevenLabs streams over WebSocket; Kokoro, then OpenAI on failure
- Per-provider usage tracked, audio cached
Adapt
Session EngineWhat Happens
- Practice mode offers adaptive hints
- Exam mode runs a 20-question certification
- Interview mode simulates the real thing
- Passing an exam issues a 1-year certificate
Four Pillars of the Voice Stack
Each pillar closes a failure class of production voice AI: availability, expressiveness, identity, and operability.
Resilient Failover
AvailabilityControls
- ElevenLabs primary — highest quality, WebSocket streaming
- Kokoro secondary — self-hosted, zero cold start, model preloading
- OpenAI tertiary — most reliable last resort
- Continuous health monitoring on the self-hosted tier
Emotion Engine
ExpressivenessControls
- Five-state EmotionContext enum
- Per-state stability, similarity, and style settings
- SUPPORTIVE and ENCOURAGING for struggle moments
- ENTHUSIASTIC and EXTREMELY_EMOTIONAL for wins
Mentor Voice System
IdentityControls
- 13 voices with structured gender, age, energy, style traits
- Per-voice Kokoro fallback IDs — persona survives failover
- Voice catalog served from a dedicated API
- #13: "Luca Sensei", the founder's own wise, calm systems-architect persona
Observability & Ops
OperationsControls
- TTS metrics exporter and per-provider usage tracking
- Streaming, generation, and download narration endpoints
- Audio cache management API
- Sentry across the whole platform
From Beta to Voice-First Platform
The full voice loop is live in beta. Each phase deepens intelligence and reach without compromising the failover-first foundation.
Beta
Live TodayDelivered
- Full voice loop with 3-tier TTS failover
- Practice, Exam, and Interview modes across 4 domains
- 8-provider BYOK LLM layer with circuit breakers
- Stripe billing and 1-year certificates
Smarter Retrieval
In DesignPlanned Deliverables
- Vector-search retrieval to replace keyword relevance
- Embeddings over the curated knowledge base
- Expanded document coverage per domain
- Retrieval-quality evaluation harness
Deeper Voice
ExplorationPlanned Deliverables
- Additional mentor personas beyond the current 13
- Latency tuning across the failover chain
- Richer per-emotion tuning from session feedback
- Expanded narration coverage for onboarding and lessons
Teams & Hiring
ExplorationPlanned Deliverables
- Organization accounts for engineering teams
- Certification as a screening signal for hiring
- Team dashboards over exam results
- New domains beyond the initial four tracks
Who Wins with Stack Sensei
Learners get a mentor who talks back; teams get verifiable skills; builders get a reference voice architecture.
For Job Seekers
For Engineering Teams
For AI Builders
Ready to Build Voice-Native AI?
This is the voice engineering I bring to client work: failover chains that survive vendor outages, synthesis that carries emotion, and LLM layers that never lock you in. If your product needs a voice — or your interview pipeline needs a sensei — let's talk.