Case Study · Sales enablement
A dedicated manager for every rep, on every call.
Manual call review is sampling — a manager spends forty minutes to evaluate one call, once, and the reps who most need coaching are the least likely to have theirs pulled. SentiVox grades every call against the campaign's own rubric and returns specific coaching the same shift. DBAI designed, built and operates the platform as its own venture.
- Project
- SentiVox
- Industry
- Sales enablement · voice AI
- Context
- A multi-tenant SaaS platform built and operated by DBAI, live at sentivox.app, running against production call volume for teams in three countries.
The bottleneck
The same three problems show up on every sales floor, and each one makes the other two worse.
- Coaching is a sample, not a practice
- Reviewing a forty-minute call takes forty minutes plus notes, so only a fraction of calls are ever heard. Feedback arrives late, unevenly, and rarely for the reps who need it most — skills develop by accident rather than by design.
- Managers are buried in work only they can stop doing
- Replay hours, note-taking, score-keeping and report assembly crowd out the one thing a manager is actually for: working directly with their reps.
- Good leads die on one bad call
- When leads are bought, a valuable one can vanish because a single conversation went badly and nobody reviewed it. A missed follow-up that nobody surfaces is indistinguishable from a lead that was never worth anything.
Before
- Review coverage
- Whatever a manager had time to listen to — a sample, chosen by availability rather than by need.
- Feedback timing
- Sporadic and delayed, arriving long after the conversation it referred to.
- Grading consistency
- Each reviewer's own judgement, applied to whichever calls they happened to pull.
- Follow-ups
- Surfaced only if someone noticed. An unworked follow-up left no trace.
What was built
A multi-tenant platform where calls arrive from the tools a team already records into, get transcribed and graded against that campaign's own weighted rubric, and land as coaching anchored to real quotes from the rep's own conversation — the same shift. The same pass works the pipeline: every call is tagged with a disposition and a lead-quality rating, so missed follow-ups surface instead of disappearing.
How it's built
01
Calls arrive without changing how anyone works
Ingestion runs on a schedule from the cloud drives a team already records into, and direct upload works standalone. Reps keep recording exactly as they did; nothing about their day changes.
02
The campaign, not the platform, decides what good looks like
Each campaign carries its own profile — weighted grading pillars, a target-persona brief, agenda notes and keywords — injected into every analysis. Changing how a campaign is graded is a form edit, not a prompt-engineering ticket.
03
Speaker-aware grading, so the right person is scored
Diarization separates who said what, so the rep is graded and coached while the customer side is read for intent rather than marked against them.
04
Coaching that accumulates per rep
Messages persist against the representative rather than the call, so each conversation adds to a development record instead of arriving as an isolated note.
05
Tenant isolation enforced in more than one place
Five roles, route-prefix middleware and per-query company scoping, so a request is checked at the edge and again in the query that serves it. Media sits in a private bucket reached only through short-lived presigned links.
Agents & automation
- Grading and coaching
- Claude grades each call against the campaign's rubric and writes coaching anchored to quotes and clips from that conversation. The model is admin-switchable from a settings panel — the identifier is stored and resolved at runtime, so moving between models is a dropdown rather than a deploy.
- Lead recovery and routing
- The same pass tags a disposition and a one-to-five lead-quality rating, surfaces missed follow-ups to a manager exception view, and hands the rating onward to the CRM.
- Scheduled ingestion and retention
- Cron jobs pull new recordings from connected drives, enforce storage caps, and apply the audio-retention policy — a call can be used once and never kept, if that is what the company chose.
- The codebase audits itself
- A findings catalogue with severity and remediation status is re-checked daily, with email alerts on regressions. On every production build the deploying diff is sent to Claude for a security review, and new findings land in that same dashboard for a human to rule on.
The stack
- CLAUDE
- Grading engine and coaching, admin-switchable at runtime
- NEXT.JS
- Application framework, App Router
- NEON POSTGRES
- Multi-tenant schema: tenancy, campaigns, calls, pillars, quotas, audit runs
- AUTH.JS
- Google sign-in, invite-gated teams and self-serve trials
- CLOUDFLARE R2
- Private call-audio storage behind presigned links
- RESEND
- Invitations and security-regression alerts
- SENTRY
- Server and client error monitoring
- VERCEL
- Hosting and scheduled jobs
Built with Claude at every layer
SentiVox was designed and built end-to-end in Claude Code, and Claude is the product's own engine as well as its build tool — it grades the calls, writes the coaching, and reviews the codebase's own diffs on every production build. That last part is the pattern DBAI carries into client work: AI in the loop, humans on the decisions, and the findings surfaced where somebody has to act on them.
In production
The platform is live at sentivox.app and running against real call volume. The figures below are operating characteristics of the system rather than claims about what it did to anyone's numbers.
- Review coverage
- Every uploaded call is graded and coached, rather than the fraction a manager had time to reach.
- Feedback timing
- Coaching lands the same shift as the conversation it refers to, anchored to quotes from that call.
- Grading consistency
- One profile per campaign, applied identically to every call on it — weighted pillars set in a form, not re-decided per reviewer.
- Follow-ups
- Every call is tagged with a disposition and a lead-quality rating, and missed follow-ups surface to a manager exception view instead of leaving no trace.
- Volume
- One deployment feeds roughly 120 calls a day through the platform, coaching SDRs across three countries.
- Deployments
- Three published deployments run on separate campaign profiles — an international SDR operation, a casino hotel and resort, and a B2B technology provider — each graded against its own rubric rather than a shared one.
Related capabilities
Every call graded, every rep coached.
Bring us the process where the review is a sample and the sample is the problem. We will show you what full coverage costs to run.
Book a Discovery Call