

THE TAKEAWAYThe cheapest option on the sheet was the one I rejected — it only spoke English, and the customers did not. The second-cheapest still cut the bill by 83%.
A five-agent lead-qualification calling system for 200 Instagram leads a day — and the hosting and scheduling decisions that costed voice at ₹45,400 a month instead of ₹2.74 lakh.
Mockingbird is a lead-qualification calling agent with a CRM attached. The client ran Instagram ads and got at least two hundred leads a day — more than the sales desk could ring before they went cold. So the agent rings them instead. A lead lands, the agent calls within minutes, qualifies them, follows up on WhatsApp if they do not pick up, and hands the sales team a warm buyer with the whole conversation attached. Worth saying plainly: this is not cold calling. Every number it dialled belonged to someone who had just filled in a form on an ad. The job was speed, not volume.
What this is, and what it is not. I built the architecture, the front end, the back end and the database, and I costed the infrastructure. I left the company before the project was finished — parts of it were still pending when I walked away. So read the numbers below as what they are: measured where I measured them, modelled where I modelled them. The call concurrency and the call durations come from testing. The monthly costs are a projection at the client's target volume, not a year of invoices. I am not claiming this ran at two hundred calls a day for months, because it did not. The engineering decisions are the real content here, and those happened.
The system: five agents, scoped on purpose. A core runtime answering as an OpenAI-compatible chat brain, a FastAPI backend routing between it and a fleet of five agents over a Docker network — the caller, WhatsApp, memory, follow-up and an analyst. A gateway orchestrator exposed eighteen CRM tools to them, scoped per agent rather than shared: memory and analyst got read-only access and could not write to a lead record no matter what they were asked to do. That scoping is not ceremony. An agent that can both summarise a conversation and edit the record it summarised will eventually edit the record to match its summary.
The telephony switch, which I did not see coming. The first build ran on Twilio, because it is the fastest way to get a number and start dialling. It worked, and then it did not: the number Twilio gave me was Canadian, and every person it rang was in India. An unknown foreign number calling an Indian mobile about an interior-design enquiry gets cut, not answered. So the stack moved to Exotel — Indian numbers, Indian carrier relationships, built for exactly this. The pickup problem was a phone-number problem, not a model problem, and no amount of prompt work would have fixed it. Exotel also gave the thing its actual shape: one number could trigger twenty calls at once. Twenty leads dialled in parallel, each call running six to seven minutes, the whole batch done inside eight. Two hundred leads clear in about an hour and a half — not two hundred calls queued one behind the other.
Then the hosting question, which was the expensive one. Two hundred calls a day, thirty days a month, six and a half minutes average — the agent is patient and does not cut people off, so its calls run longer than a human's. That is thirty-nine thousand minutes a month, six thousand calls. Managed voice platforms bill by the minute. At that volume the minute is the whole business model, so I costed three routes against the same requirement before anyone committed to anything.
Route one — the managed voice platform already in use. Its top tier was ₹87,120 a month for 12,375 included minutes: ₹7.04 a minute, billed for the whole month whether the campaign runs that long or not. Thirty-nine thousand minutes is more than three times what that plan covers. The rest bills as overage at a rate the vendor would not confirm in writing. Taken at the plan's own effective rate: ₹2,74,560 a month, ₹45.76 a call, for voice alone. Concurrency was also capped by tier — which matters when twenty calls go out at once rather than one at a time.
Route two — a pay-per-minute voice agent platform. No subscription, billed on usage, with telephony and number provisioning included, which the self-hosted route does not have. But the headline per-minute rate only covers the voice engine. Add the language model and the telephony leg and a realistic production rate lands near ₹12.35 a minute — about ₹4,81,650 a month at our volume. The option marketed as the cheap one was the most expensive on the sheet, at exactly the scale we needed.
Route three — self-hosted, and the one I recommended. No subscription, no per-minute fee, no call-duration cap. Pay for the GPU by the hour and nothing else. The honest caveat went into the document in the same breath: a compute figure is compute only. It carries no phone number and no telephony, both of which the managed routes bundle in. It was never an apples-to-apples number and I did not present it as one — the Exotel leg sits on top of it. It also needed roughly two weeks of build and training before it could place a single call.
And then the cheapest option on the sheet got rejected — by me, on the one ground that mattered. The self-hosted voice stack we had costed spoke English. The client's customers speak Marathi, Hindi and English, often inside the same sentence. Bending an English-only model into a trilingual one is not a configuration change; it is a different project with a worse result. So the architecture stayed self-hosted and the audio layer was swapped for an open-weight multimodal speech model — speech in, speech out, all three languages in one voice, no separate engine per language — with its API wired straight into the calling agent. A model that is cheaper and speaks the wrong language is not cheaper. It just fails for free.
What ran, and where — which is the part I would defend hardest. An NVIDIA A100 with 80 GB of VRAM and 128 GB of system RAM, at roughly ₹144 an hour, carried only the things that genuinely need a GPU: the speech model holding the live conversation, Whisper transcribing each call for the record the CRM keeps, and the language model behind the caller. That language model was Qwen2.5-7B, fine-tuned on real sales material rather than prompted into behaving. It had exactly one job — being the voice on a call — and it did nothing else. Mockingbird itself, the orchestrator the whole system is named after, ran on Gemini's free tier. Deciding when to bring the GPU up, which leads are due, when a retry round should go out, what the desk should see in the morning: low volume, no latency budget, no reason to pay for it. Its billing never crossed ₹250 in a month. That split is the entire cost argument in miniature. The model that bills ₹144 an hour only runs while somebody is actually on the line. The model that decides when that should happen is free.
Twenty concurrent calls was the design target, because twenty is roughly what one A100 holds. Concurrent language-model streams do not cost twenty times one stream — the card batches them and reads its weights once, so twenty calls are closer to one and a half times the load than twenty. But the latency budget on a live call is unforgiving. Twenty was the ceiling, not comfortable headroom, and anyone who tells you a single card scales indefinitely has not watched a call go quiet while a model thinks. Everything that was not a model sat on a separate VM at around ₹2,000 a month: the five agents, the gateway orchestrator, the CRM, the database and the front end. None of it needs a GPU, and pinning it to the same box would have meant paying GPU rates to run a web app.
The last saving was a schedule, and the agent owned it. A GPU billed by the hour is a machine idle two-thirds of its life, because nobody answers a sales call at four in the morning. So the dialling window became infrastructure. I gave Mockingbird scoped CLI access to the Jarvislabs account and let it run its own server: up at 10:00, down at 20:00, the only ten hours the agent is allowed to dial anyway. It also checked before it spent anything. No new leads that day meant it did not start the calling agent at all. New leads arriving through the afternoon triggered a fresh round. And every number that did not pick up came back around three hours later for another attempt — a retry pass, scheduled rather than hoped for, and a meaningful share of the pickups came from exactly there. An agent deciding when its own GPU is worth running is the part I would build again.
The numbers, end to end. 10 hours × 30 days = 300 GPU hours a month at ₹144 — ₹43,200. With the application VM and the free-tier model billing on top, about ₹45,400 a month, or roughly ₹7.57 a call. Against ₹2,74,560 on the managed plan, that is ₹2,29,160 saved a month — close to ₹27.5 lakh a year, an 83% cut, and about six times cheaper per call. The window is worth pricing on its own. The same card left running around the clock would have cost ₹1,03,680 a month instead of ₹43,200 — ₹60,480 a month for hours nobody would have dialled in. That is not part of the 83%; it is a separate saving, and it came from treating the schedule as architecture instead of an afterthought.
The behaviour came from studying one practitioner rather than writing prompts — Saad Khaja, who posts as @saadsells. Not his slogans; his mechanics. How an objection gets acknowledged before it gets answered, what a second follow-up says that a first one cannot, when a call should be ended rather than saved. His published material was the behavioural reference the agent was designed against. Credit where it belongs: the agent sounds like it knows what it is doing because somebody who does was studied closely. The method is his.
The demo linked below is the real front end, running on mock data. The backend is deliberately not part of it — the client's system is theirs, and a portfolio piece has no business carrying it. Everything on screen is the interface as built: the lead desk, the conversation threads, the call log, the qualification scoring. Sign in with any email. The prefix picks the role — admin@, sales@ or viewer@.
The promo film — how a lead becomes a booked site visit
Open the live demo — real front end, mock data
mockingbird.wishmaster.space