How Apna scaled 75 Lakh AI interview minutes using BlueMachines.ai and ElevenLabs.ai

In Apna’s AI Interview Prep, this stack powers every moment — from when a candidate says “Tell me about yourself” to when an AI interviewer responds empathetically with “That’s interesting — could you explain how you handled that project?”.

Each turn triggers real-time orchestration of speech-to-text, language understanding, context retrieval, and voice synthesis, all synchronized seamlessly through BlueMachines.

It’s the same backbone that powers enterprise-grade deployments of BlueMachines.ai.

Under the hood, BlueMachines.ai runs on a six-layer architecture engineered for precision, compliance, and resilience — designed not just to run voice AI, but to orchestrate conversation itself in under 300 milliseconds.

1. Connectivity & Compliance Layer

Regional edge gateways ensure all sessions stay low-latency and compliant with Indian TRAI and RBI norms.

Apna’s voice data never leaves the region — essential for BFSI and education partners that demand full data residency.

2. Speech Processing Layer

Every interview begins with a candidate’s voice — captured, cleaned, and transformed in real time.

The Speech Processing layer converts raw audio into precise text using multilingual ASR capable of understanding English, Hindi, and regional dialects, often mixed in the same sentence (e.g., “Main ne growth 25% tak increase kiya last quarter”).

Key features include:

• Noise and echo cancellation for crowded environments where even the “fan” noise gets removed to ensure accurate downstream processing.

• Voice Activity Detection (VAD) to distinguish speech from silence.

• Adaptive Code-Switching Models for fluid bilingual conversations between Hindi <> English <> Hinglish.

Each spoken phrase reaches the language model in ~100 ms, preserving the natural rhythm of speech.

3. Conversational Intelligence (RAG and Workflow Layer)

Once text is available, BlueMachines routes it into the Conversational Intelligence layer — where meaning, tone, and intent take shape.

This layer blends RAG-based retrieval, Persona logic and User Workflows to decide the next move in the conversation.

For Apna’s AI Interview Prep, this means dynamically generating role-specific and company-specific questions based on a candidate’s resume and the targeted job description.

For instance, a candidate interviewing for a Sales Manager role at HDFC might face questions around conversion ratios, while one preparing for a Data Analyst position at Flipkart gets probed on SQL joins and funnel optimization.

The process is near-instantaneous:

1. The candidate’s latest response is parsed and classified.

2. The RAG engine fetches contextual data from a pre-indexed library of 500M role–company patterns.

3. The LLM ensemble (OpenAI, Anthropic, or in-house tuned models) composes the next question, calibrated for difficulty and tone.

All of this happens in parallel threads — returning output to the Orchestration Core within ~120 ms.

4. Security & Reliability

BlueMachines ensures AES-256 encryption, RBAC controls, and zero-downtime failover.

If an AI node drops mid-interview, session state transfers instantly to a sibling pod with no perceptible pause.

Every turn is logged through the observability mesh for latency, routing accuracy, and contextual drift — all while remaining compliant with ISO 27001 and SOC 2 standards.

5. Orchestration Core — The Heartbeat of the System

The Orchestration Core is where BlueMachines’ engineering truly shines.

It coordinates every millisecond of traffic between the STT, LLM, and TTS pipelines — the three pillars of real-time conversation.

STT (Speech-to-Text)
Receives the candidate’s voice - cleaned through VAD and echo suppression, and streams partial transcripts to the NLU before full sentences are complete. This streaming mode minimizes delay and allows predictive intent modeling.

LLM (Language + Context Layer)
Runs the reasoning logic, retrieves relevant context from Apna’s company-role knowledge graph, and triggers persona-specific question generation. This enables realistic interviewer personas — HR, Hiring Manager, or Technical Lead — each with distinct tone, patience, and follow-up depth.

TTS (Text-to-Speech)
This is where ElevenLabs.ai transforms machine responses into human voice with remarkable fidelity.

ElevenLabs’ multilingual, emotionally nuanced voices are tuned for Indian English, Hindi, and code-mixed speech, allowing the AI interviewer to modulate tone — empathetic when encouraging, firm when challenging.

Each synthesized response begins playback within ~150–180 ms, thanks to ElevenLabs’ low-latency streaming APIs integrated directly into BlueMachines’ orchestration layer.

Combined, these three layers create a continuous conversational loop — fast enough that users forget they’re talking to AI.

“Every millisecond is a design choice. We don’t just monitor latency — we architect it,” says Abhishek Ranjan, CTO, BlueMachines.ai.

6. Continuity & Governance

The final layer ensures explainability and consistency.

Each session is auditable, replayable, and testable through nightly synthetic interviews that catch prompt drift or bias before production.

Every score is logged, replayable, and auditable — meeting enterprise compliance without sacrificing latency.

Persistent memory allows the AI to remember candidate progress and resume from prior sessions, enabling a continuous, personalized learning journey.

Every millisecond is designed, not assumed.

Every voice packet is a proof of intent — that infrastructure can feel human.