Challenge
LeadTrack AI, a US sales-tech company, came to us with a problem their customers felt every day: inbound interest has a half-life of minutes. A prospect who fills in a form at lunch is effectively cold by dinner, which means response time — not product quality — decides who wins the deal. No human sales team can answer every lead within seconds, around the clock, forever.
Four failure points defined the brief:
- Coverage gaps. Nights, weekends and campaign traffic spikes went unanswered entirely.
- Inconsistent qualification. Every rep asked different questions in a different order, so pipeline data was never comparable.
- No context handoff. When a hot lead did reach a human, it arrived as a bare phone number with no history attached.
- Speed as a staffing problem. The only known fix was hiring more people in more time zones — expensive, and it still capped out.
The engineering challenge was harder than the commercial one. Prospects hang up on anything that sounds like a phone tree, so the voice agent had to handle interruptions, tangents and varied accents in real time. Voice AI also tolerates no awkward pause — every millisecond of pipeline latency shows up as dead air on a live call, and dead air ends calls. On top of that, this had to ship as a multi-tenant SaaS product, not a single deployment: every customer needed their own agents, scripts, phone numbers, pipelines and billing, with call data hard-isolated between tenants while still sharing one codebase and one improvement loop
LeadTrack AI, a US sales-tech company, came to us with a problem their customers felt every day: inbound interest has a half-life of minutes. A prospect who fills in a form at lunch is effectively cold by dinner, which means response time — not product quality — decides who wins the deal. No human sales team can answer every lead within seconds, around the clock, forever.
Four failure points defined the brief:
- Coverage gaps. Nights, weekends and campaign traffic spikes went unanswered entirely.
- Inconsistent qualification. Every rep asked different questions in a different order, so pipeline data was never comparable.
- No context handoff. When a hot lead did reach a human, it arrived as a bare phone number with no history attached.
- Speed as a staffing problem. The only known fix was hiring more people in more time zones — expensive, and it still capped out.
The engineering challenge was harder than the commercial one. Prospects hang up on anything that sounds like a phone tree, so the voice agent had to handle interruptions, tangents and varied accents in real time. Voice AI also tolerates no awkward pause — every millisecond of pipeline latency shows up as dead air on a live call, and dead air ends calls. On top of that, this had to ship as a multi-tenant SaaS product, not a single deployment: every customer needed their own agents, scripts, phone numbers, pipelines and billing, with call data hard-isolated between tenants while still sharing one codebase and one improvement loop
Solution
We built LeadTrack AI as a multi-tenant SaaS platform that turns speed-to-lead from a staffing problem into an infrastructure property — an AI sales floor that never sleeps.
What we built
- AI voice agents that qualify through natural, interruptible conversation rather than a scripted survey — the agent handles being cut off, follows tangents, and copes with real-world accents and line quality.
- Instant auto-dialling that calls every new lead within thirty seconds, at any hour, so interest is met while it is still warm.
- Per-tenant qualification logic — each customer configures their own scripts, questions and intent-scoring model, mapped to how they actually define a qualified opportunity.
- Context-rich human handoff — high-intent prospects are routed live to a rep who joins with the full transcript and an intent score already in hand, so the conversation continues instead of restarting.
- A true multi-tenant platform with isolated tenant data, per-customer configuration, numbers and billing on a single codebase.
- Call analytics and behavioural evaluations across every conversation, so outcomes are tracked and agent behaviour keeps improving rather than drifting.
How we delivered it
We worked in four phases. In conceptualisation, we modelled real qualification conversations and escalation rules alongside working sales teams. Design produced the tenant dashboards for scripts, pipelines, transcripts and outcomes. Development covered the voice-agent pipeline, telephony integration and the multi-tenant spine. Deployment hardened the system against live production traffic rather than a test set.
Stack: Node.js and NestJS on the application layer, Python for the AI and voice pipeline, MySQL for transactional data, Redis for low-latency state, ElasticSearch for transcript and call search, all running on AWS with a DevOps setup built for telephony-grade uptime.
We built LeadTrack AI as a multi-tenant SaaS platform that turns speed-to-lead from a staffing problem into an infrastructure property — an AI sales floor that never sleeps.
What we built
- AI voice agents that qualify through natural, interruptible conversation rather than a scripted survey — the agent handles being cut off, follows tangents, and copes with real-world accents and line quality.
- Instant auto-dialling that calls every new lead within thirty seconds, at any hour, so interest is met while it is still warm.
- Per-tenant qualification logic — each customer configures their own scripts, questions and intent-scoring model, mapped to how they actually define a qualified opportunity.
- Context-rich human handoff — high-intent prospects are routed live to a rep who joins with the full transcript and an intent score already in hand, so the conversation continues instead of restarting.
- A true multi-tenant platform with isolated tenant data, per-customer configuration, numbers and billing on a single codebase.
- Call analytics and behavioural evaluations across every conversation, so outcomes are tracked and agent behaviour keeps improving rather than drifting.
How we delivered it
We worked in four phases. In conceptualisation, we modelled real qualification conversations and escalation rules alongside working sales teams. Design produced the tenant dashboards for scripts, pipelines, transcripts and outcomes. Development covered the voice-agent pipeline, telephony integration and the multi-tenant spine. Deployment hardened the system against live production traffic rather than a test set.
Stack: Node.js and NestJS on the application layer, Python for the AI and voice pipeline, MySQL for transactional data, Redis for low-latency state, ElasticSearch for transcript and call search, all running on AWS with a DevOps setup built for telephony-grade uptime.
Results
The platform is live and running at production volume, and the numbers hold at scale rather than in a pilot.
Under 30 seconds to first call. Every lead is engaged inside the window where interest is still warm — nights, weekends and campaign spikes included. Coverage stopped being a rota problem.
100,000+ calls completed. This is the number that matters most technically: the agents were hardened against a hundred thousand real conversations, with real interruptions, real accents and real objections, not against a curated test set.
38% lift in conversion. More qualified conversations, reached sooner, with humans spending their time only on the calls that had earned the attention.
Beyond the headline metrics, qualification became consistent and measurable for the first time — every tenant now scores leads against the same defined criteria, so pipeline data is comparable across reps and campaigns. Sales teams start conversations warm, with a transcript and intent score instead of a blank phone number. And because the platform is multi-tenant, LeadTrack AI can onboard a new customer onto its own isolated agents, scripts and pipelines without a new deployment.
As the client's founder, Neel Bhattacharya, put it: "I've worked with three agencies before CODT. The difference: they pushed back on the spec when it didn't make sense, and shipped what we actually needed."
Two things worth checking before you submit: Techreviewer usually strips markdown, so the bold and bullets may need converting to plain lines or dashes. And if the testimonial line pushes you over a limit or feels out of place in a "Results" field, cutting the final paragraph still leaves you well over the 200-character minimum.
The platform is live and running at production volume, and the numbers hold at scale rather than in a pilot.
Under 30 seconds to first call. Every lead is engaged inside the window where interest is still warm — nights, weekends and campaign spikes included. Coverage stopped being a rota problem.
100,000+ calls completed. This is the number that matters most technically: the agents were hardened against a hundred thousand real conversations, with real interruptions, real accents and real objections, not against a curated test set.
38% lift in conversion. More qualified conversations, reached sooner, with humans spending their time only on the calls that had earned the attention.
Beyond the headline metrics, qualification became consistent and measurable for the first time — every tenant now scores leads against the same defined criteria, so pipeline data is comparable across reps and campaigns. Sales teams start conversations warm, with a transcript and intent score instead of a blank phone number. And because the platform is multi-tenant, LeadTrack AI can onboard a new customer onto its own isolated agents, scripts and pipelines without a new deployment.
As the client's founder, Neel Bhattacharya, put it: "I've worked with three agencies before CODT. The difference: they pushed back on the spec when it didn't make sense, and shipped what we actually needed."
Two things worth checking before you submit: Techreviewer usually strips markdown, so the bold and bullets may need converting to plain lines or dashes. And if the testimonial line pushes you over a limit or feels out of place in a "Results" field, cutting the final paragraph still leaves you well over the 200-character minimum.
