How AppUnik Built SalesEar's AI Call Intelligence Platform on Claude API

Field sales teams in India have a logging problem. After every call, agents are supposed to update the CRM with what was discussed, what the prospect needs, and what the next step is. In practice, most agents skip it. The ones who do log something write "prospect interested, will follow up" and move on to the next call.

The result: deal context disappears the moment the call ends. Managers run pipeline reviews based on incomplete notes. Follow-ups get missed because nobody remembers what was promised. When a lead goes cold, nobody can point to the specific call where it happened.

Kejal Dave, Co-Founder of SalesEar, had spent enough time around Indian sales operations to know this was not a discipline problem. It was a workflow problem. Asking agents to stop after every call and type a detailed summary is asking them to do something that actively slows them down. The solution was not better CRM training. It was removing the logging step entirely and replacing it with AI that listens to the call and generates the summary automatically.

She had the domain expertise and the product vision. What she needed was a technical team that could build it.

The Engineering Challenge

The product concept sounds straightforward: record sales calls, transcribe them, and generate useful insights. In practice, it is anything but.

Indian sales calls are multilingual. An agent in Ahmedabad pitches in Gujarati, switches to Hindi when discussing pricing, and drops English financial terms throughout. An agent in Kolkata does the same with Bengali, Hindi, and English. Standard speech-to-text tools trained on monolingual datasets produce garbage on this audio.

The call recordings come from personal Android phones, not VoIP systems. Samsung, Xiaomi, and Google Dialer devices each handle call recording differently at the OS level. Capturing audio reliably across this device fragmentation is a non-trivial mobile engineering problem.

And the AI analysis layer needs to do more than summarize. It needs to detect pricing commitments, flag compliance risks, score agent performance, identify objection patterns, and track how a prospect's intent shifts across multiple calls. This is not a prompt-and-response chatbot. It is a structured intelligence extraction pipeline that needs to produce consistent, reliable output on every call.

Kejal brought this problem to AppUnik.

Why Claude API

AppUnik evaluated multiple LLM options for the analysis engine. The decision came down to what the model needed to do with sales call transcripts: extract structured data reliably, handle multilingual input without losing context, and produce output that could feed directly into a scoring and alerting system.

Claude API, specifically Sonnet 4.6, delivered on what mattered for this use case.

The analysis pipeline does not need creative writing. It needs the same structured fields extracted from every call: summary, score, objections detected, commitments made, next actions, intent level, and coaching signals. Claude's instruction-following produced consistent structured output across thousands of calls without the format drift that other models exhibited on long-running extraction tasks. When you are processing calls at volume, you cannot afford a model that returns a different JSON shape on call 500 than it did on call 1.

Then there is the multilingual problem. A sales call transcript where the agent says "sir, property ka rate 85 lakh hai, negotiable nahi hai, parking included hai" contains pricing information in three languages within one sentence. The analysis model needs to understand that 85 lakh is a price, "negotiable nahi hai" is a firm commitment, and "parking included" is an amenity promise, regardless of which language each phrase is in. Claude handled this cross-lingual reasoning without requiring separate language-specific processing steps.

The third piece was conversation understanding. Sales calls are not transactional. They have objections, deflections, rapport-building, and subtext. An agent who says "main manager se baat karke batata hoon" is not making a commitment to call back. They are deflecting. Claude's ability to distinguish between genuine commitments and polite deflections was measurably better than alternatives on the Indian sales call transcripts AppUnik tested during evaluation.

The Build

AppUnik handled the full technical delivery: architecture, backend, frontend, mobile, AI pipeline, and production deployment.

The backend runs on NestJS with PostgreSQL, handling call ingestion, user management, organization hierarchies, and the analytics pipeline. The AI analysis layer processes calls through a two-tier system: a fast standard tier for high-volume processing and a deep tier powered by Claude for calls that need higher accuracy on multilingual content, compliance detection, and detailed coaching insights.

The mobile layer captures calls from the agent's phone without requiring a separate dialer app. On Samsung and Xiaomi devices, recordings are picked up automatically from the built-in call recorder. On Google Dialer devices, a lightweight helper application automates the capture. The agent's workflow does not change.

The frontend is a Next.js dashboard where managers see every call scored, transcribed, and analyzed. Call journey views group every interaction with a prospect across all agents. Persona match profiles classify buyers based on their behavior across multiple calls. Follow-up gaps, pricing flags, and commitment tracking surface automatically.

Transactional email runs through the salesear.com domain via Hostinger SMTP. The entire system is production-deployed and serving real customers.

What Changed

Before SalesEar, the sales teams Kejal worked with had one version of reality: whatever agents chose to type into the CRM. After SalesEar, they have a complete version: every call transcribed, every commitment recorded, every follow-up gap visible.

The measured impact after deploying to the first customer teams:

Call logging coverage went from roughly 30% (agents manually updating the CRM after some calls, skipping most) to 100% automated coverage on every call. Managers who previously worked from incomplete notes on 3 out of 10 calls now have a structured transcript, quality score, objection flags, and recommended next actions for every conversation. No agent behavior change was required to achieve this because the capture is automatic.

Daily call review time dropped from approximately 60 minutes of listening to recordings to under 15 minutes of reading flagged transcripts. Managers stopped sampling random calls and started reviewing only the calls with pricing flags, compliance concerns, or follow-up gaps. The review process became targeted instead of random.

Follow-up gaps that were previously invisible became visible within 48 hours of deployment. Agents who logged callbacks in the CRM without actually making the call were identified immediately because the system tracks actual calls, not self-reported activity. In the first week, managers discovered that over 30% of logged follow-ups had never actually happened.

Claude's structured output consistency was central to these results. Every call produces the same set of fields: summary, score, objections, commitments, next actions, intent level. This consistency means the dashboard data is reliable across hundreds of calls per day. Managers trust the data because it does not vary based on which agent logged it or whether they remembered to log it at all.

The business traction followed the product results. Production launch completed and stable. One customer on an annual commitment. Six new signups in a single week from organic discovery, including AI search referrals where ChatGPT recommended SalesEar to a prospect unprompted. An active conversion pipeline with enterprise-level prospects, including a multi-billion-rupee real estate developer evaluating the platform for a 50 to 200 agent deployment.

"AppUnik turned our vision for SalesEar into a working product faster than we expected. Their Claude API expertise meant we had AI-powered call analysis in production, not just in a demo." Kejal Dave, Co-Founder, SalesEar

The Technical Partnership Model

AppUnik operates as a full-stack technical delivery partner for AI-first products. The SalesEar engagement covered the complete lifecycle: product architecture, AI pipeline design and integration, backend and frontend development, mobile engineering, DevOps, and production support.

For founders with deep domain expertise but no engineering team, this model compresses the path from concept to production. Kejal brought the sales operations knowledge and the product vision. AppUnik brought the engineering capability and the Claude API integration expertise. SalesEar went from concept to production-grade platform serving real customers.

The Claude API integration is not a wrapper around a chatbot. It is a production pipeline that processes thousands of multilingual sales calls, extracts structured intelligence, and feeds it into a scoring, coaching, and compliance system. Building this required understanding both the AI capabilities and the sales domain constraints. AppUnik bridged both.

AppUnik is an AI-first software company based in Ahmedabad, specializing in building production AI systems using Anthropic's Claude API. For technical partnerships and AI product development inquiries, contact info@appunik.com.