- Platform
- Web App
- Duration
- 12 weeks
- Industry
- Marketing Tech
- Read time
- 4 min read
In short
RaftLabs built a voice-first interview platform for Perceptional that calls respondents by phone and conducts automated AI interviews. The platform uses Twilio for outbound calling and ElevenLabs for voice synthesis, delivering 6x deeper insights than text surveys with results available as soon as the last call ends. Interviewers upload a contact list, the system calls each number automatically, and an AI conducts the conversation adapting questions in real time based on what the respondent says. Built and launched in 12 weeks.
Cherian Koshy runs feedback programs for large organizations across the US. For years, his platform collected responses through text surveys. The technology worked fine. The data was the problem.
The people hardest to reach, including frontline workers, older staff, and anyone who found typing slow or uncomfortable, were the ones least likely to complete a survey. Results skewed toward the most digitally comfortable respondents. Cherian knew it, and he wanted to fix it.
He came to us with one brief: build a system that reaches anyone with a phone.
We rebuilt Perceptional as a voice-first platform. Interviewers upload a list of phone numbers. The platform calls each one automatically. An AI conducts the interview by phone, listens to what the person says, and selects the next question based on the response. When the last call ends, the analysis is ready. No web access required. No typing. No links to click.
before / after
What changed
- Only reached people comfortable with typing and filling out online forms
- High friction: recipients had to click a link, enter personal details, and type out answers
- No way to capture tone, hesitation, or emotion: only what people wrote
- Reports arrived days after the survey window closed, often after the team had moved on
- No retry system for missed respondents without running another email campaign
- Basic reporting with no sentiment tracking or theme detection
- Anyone with a phone can participate: no web access, account, or typing required
- The platform calls recipients directly and the AI conducts the interview by phone
- Conversations adapt in real time to what the person says, the way a human interviewer would
- Reports are ready as soon as the last call ends
- Missed calls trigger an automatic retry or SMS fallback: nothing falls through the cracks
- Sentiment analysis, keyword tracking, and feature reports come included
What we had to solve
- 01
Running hundreds of calls at the same time
Large feedback campaigns send 500 or more calls in a single batch. Phone infrastructure does not behave like a web app. Each call has its own connection state, audio variables, and carrier-side failure modes. We built retry logic that tells the difference between a busy signal, a network drop, and a declined call, then handles each one differently without human involvement. The platform runs hundreds of simultaneous calls without a single one blocking another.
- 02
Getting people to stay on the call
Early test calls had a real problem: people hung up. The AI sounded robotic at the joins between responses, the pacing felt slightly off, and the conversation was too obviously scripted. We ran multiple rounds of ElevenLabs voice tuning and prompt work until the dropout rate matched what a human interviewer running the same questions would expect. The final version does not sound like an automated phone tree.
outcomes
What we achieved
The original platform was text-only. Cherian came to us with a prototype idea, not a working product.
Text surveys returned surface-level answers from the most digitally comfortable respondents. The data was useful but incomplete.
Research teams waited days after survey windows closed before they could access and analyze the data.
Running feedback programs that miss the people hardest to reach?
the build
What we built
The build came down to three things: reach anyone with a phone, make the conversation feel real, and give teams data they can act on the same day.
Recipients pick up and speak normally, no app, no link, no account to create
The interviewer uploads a CSV of phone numbers. The platform calls each one automatically. The recipient picks up and speaks normally. No account to create, no link to visit, no personal details to enter. The AI asks questions in order, listens to the response, and picks the follow-up that fits best. A full transcript is saved and ready for analysis when the call ends.
Interviewers watch the campaign run start to finish without touching individual records
A dashboard shows every call in real time: completed, failed, pending. For missed calls, one click retries the call or sends an SMS fallback with a web-based interview link. Completed respondents are locked so nobody gets called twice. The interviewer watches the campaign run from start to finish without touching individual records.
Feature reports, keyword themes, and sentiment, no manual transcript reading required
Feature reports surface what was said about specific topics. Keyword analysis finds emerging themes across all interviews. Sentiment analysis shows how people felt, not just what they said. Each report type is exportable as a PDF for sharing with the team. No manual transcript reading required.
Teams ask questions about their interview data directly, hours of manual review replaced
An AI chatbot lets the team ask questions about their interview data: which pain points came up most, what people said about a specific product feature, how sentiment varied across departments. It replaces hours of manual transcript review with a direct conversation with the data.
Engagement
How we worked together
- 01Weeks 1–2
Discovery and scoping
We map the problem before writing code. Two weeks of technical audit, stakeholder interviews, and prototype — so both teams align on scope and risk before sprint one.
- 02Ongoing
Two-week Agile sprints
Each sprint ends with working software, not a status update. You review a real build, request changes, and approve before we move forward. No surprises at handover.
- 03Ongoing
Daily async updates
Slack for daily progress, Asana for task visibility, weekly video calls for decisions. You have full visibility without needing to attend every meeting.
- 04Final
Handover and warranty
Full code handover with deployment runbooks and documentation. Thirty-day warranty period for production issues at no extra cost.
stack
Why we chose this stack
- 01Twilio handles outbound calls worldwide through a single API. No carrier agreements to negotiate, no per-country setup. A call goes out anywhere in the world with one request.Twilio
- 02The voice had to sound human, not robotic. ElevenLabs gave us the quality we needed at a price that made a bootstrapped product viable. Without it, the dropout rate would have made the platform unusable.ElevenLabs
- 03Cherian was building this solo. Lambda scales to 1,000 simultaneous interviews without an infrastructure team. No idle capacity, no ops overhead between campaigns.AWS Lambda
- 04All interview transcripts, sentiment scores, and analytics needed reliable relational storage with fast querying. PostgreSQL handles this at scale without added complexity.PostgreSQL
Still curious?
12 weeks from first conversation to live product. We built a working prototype in the first week so Cherian could see the call flow in action before committing to the full build. Two-week Agile sprints from there, with daily updates throughout.
Yes. The platform runs on AWS Lambda, which scales automatically to handle thousands of concurrent calls. The same architecture that handled the beta handles large-scale campaigns. There is no capacity ceiling to manage manually.
Voice captures what text cannot: hesitation, emphasis, emotion, correction. The AI adapts its questions based on what the person said in the previous response, so it drills down where a survey would just move on. You also reach respondents who would never complete a text form, which removes a systematic bias from your data.
The system retries automatically or switches to an SMS fallback with a web-based interview link. Each phone number is treated as unique per campaign, so nobody gets called more than once.
Research teams, HR departments, product organizations, and market research agencies: anyone collecting qualitative feedback at scale. It works equally well for a 50-person employee survey and a 2,000-person customer research campaign.
Related work
More work like this
An AI OCR loyalty platform drove 1,000+ sign-ups for a major supermarket chain
We built an AI loyalty platform for Musgrave's SuperValu and Centra that automates receipt validation with 99.9% accuracy, acquiring 1,062 users in 4 weeks.
Read case study
LoyaltyPass delivers wallet-native loyalty cards with 60% higher enrollment than app-based programs
LoyaltyPass runs loyalty programs inside Apple Wallet and Google Wallet. Customers earn and redeem without a separate app download, and businesses get real engagement data without the friction of a native app.
Read case study
Energia rewards platform logs 1,000+ logins in 24 hours after WordPress-to-Next.js rebuild
We rebuilt Energia's loyalty rewards platform from WordPress to Next.js in 12 weeks, migrating 300K+ user records, automating daily SFTP workflows, and delivering a system that hit 99.9% uptime from day one.
Read case study
