Conversation Quality Analyst

Bolna is Voice AI infrastructure built for India - and now for the world. We help businesses deploy intelligent voice agents that can call, converse, and convert in any language, at scale. From collections to customer support to sales, our agents handle millions of conversations so humans don’t have to.

We’re a YC F25 company, backed by General Catalyst, with 1,050+ paying customers and growing fast. Our team of ~25 is based in Bengaluru.

Every voice AI agent Bolna deploys makes real-time judgment calls - when to speak, when to go silent, when a customer is done talking, when to hand off. We’re building automated systems to grade these calls at scale, using LLMs as judges of call quality. But before you trust a model’s judgment, you verify it against a human’s.

That’s this role. You’ll listen to real calls, annotate what actually happened, and check whether our automated systems - LLM-as-judge evals and quantitative signal detection - got it right. It’s precise, high-attention work, and it sits right at the center of how we know our voice agents are actually working.

This is an internship role for someone early in their career who wants hands-on exposure to how a voice AI company builds trust in its own AI.

• Annotation: Listen to and annotate real customer calls - transcription review, issue tagging, labeling - using tools like Label Studio. Follow (and help sharpen) annotation guidelines for a multilingual environment (Hindi, English, Hinglish).
• Verifying LLM-as-Judge Evaluations: For calls flagged by our automated eval pipeline, verify whether the model’s call was actually correct - for example, confirming whether a detected barge-in (agent/customer talking over each other) genuinely happened by listening to the audio. Mark agreements and disagreements clearly, with reasoning, so we can measure and improve model accuracy over time.
• Verifying Quantitative Measures: Check system-flagged quantitative signals against the actual call - e.g., confirming whether an “agent interruption” the system detected really occurred at that timestamp. Flag false positives/negatives so we can tighten detection logic. Help identify edge cases that current rubrics or detection logic don’t handle well.
• Inspecting Calls & Surfacing New Issues: Regularly inspect calls beyond flagged ones to spot new or emerging issues our rubrics and detection systems don't yet cover. Bring these patterns back to the team so rubrics, prompts, and detection logic keep improving.

Must-have:

• Strong attention to detail and the patience to do focused, high-precision work across many calls.
• Multilingual comfort preferred - Telugu, Tamil, Kannada, Marathi, Gujarati, or Bengali, in addition to English/Hindi.
• Comfortable learning new tools quickly - Label Studio, dashboards, internal QA apps.
• Genuine curiosity about AI and voice AI - you want to understand why a call was flagged, not just complete a checklist.

Good to have:

• Any prior exposure to data annotation, labeling, or QA work.
• Familiarity with spreadsheets/basic SQL or comfort reading dashboards (e.g., Metabase).
• Background in linguistics, call center operations, or content moderation.

• Direct exposure to how a fast-growing AI infra company builds trust in its own models.
• Real ownership over a function (call quality) that directly affects what customers see.
• Fastest way to learn the guts of voice AI - ASR, agent logic, eval pipelines - from the ground up.
• Bengaluru office, in-person team collaboration.