The call you missed
was a customer.
Swaram is a 24×7 AI voice agent for Indian small businesses. It answers the phone, the website and the app — in Hindi, Hinglish, Odia, Tamil and more — books the appointment, takes the order, captures the lead, and hands the hard calls to a human. Start on Overview for the pitch, then play every feature for yourself.
What Swaram is
One sentence, the way we'd say it in a pitch room.
Swaram gives any Indian small business a 24×7 AI voice agent that answers the phone and the website — in the customer's own language — books appointments, captures leads, takes orders, and hands difficult conversations to a human.
Built in India, in Indian languages, for the businesses global tools ignore. Self-serve, starts free, priced for a kirana — not a call centre.
The problem, in rupees
Every missed call is a customer who called someone else. Move the sliders to your own numbers.
How it works
Three steps, and the owner does all of them from a phone.
Tell it about your business
Photograph the rate card, paste your website, or just let the agent interview you for five minutes. No forms, no typing out a catalogue.
Pick where it answers
Your phone number, a button on your website, or a link and QR code we host for you. Same agent on all three — set up once.
It starts answering
Every call picked up, in your customer's language. Bookings and leads land in your app. Anything it can't handle rings you, with the full story attached.
What it can do today
Every card here is live and demoable. Click any of them to jump straight to that demo.
Who it's for
The long tail of 2-to-50-person businesses that lose money to missed calls and can't afford a receptionist.
Clinics & labs
Appointments, timings, reports, doctor availability. Every missed call is a missed consultation.
Kirana & wholesale
Order taking, stock checks, delivery slots, payment links. Orders arrive by phone, not by app.
Coaching centres
Batch enquiries, fees, demo classes. Admission season is entirely a phone business.
Logistics & transport
Docket tracking, ETAs, quotes. High call volume, mostly the same five questions.
Salons & spas
Bookings, price lists, stylist availability. Staff are with clients, not at the phone.
Real estate
Site enquiries, availability, visit scheduling. Leads go cold within the hour.
Repair & services
Job booking, charges, technician dispatch. The owner is usually on-site, not at a desk.
Restaurants & catering
Table bookings, takeaway orders, party enquiries. Peak calls hit exactly when nobody is free.
Pricing
A bundle of minutes, not a per-minute meter. In rupees, upgradeable in one tap over UPI.
Why not just use a global tool
Ringg.ai is the closest reference point and a useful contrast — we deliberately don't fight on their turf.
Where it goes next
Be honest about the boundary. The left column is demoable today; the right is the vision, and shouldn't be sold as present tense.
Why this matters beyond the business case
Made in India, for India
Indian-language speech models, tuned for how India actually talks — code-mixed, accented, on a bad line. Global tools are English-first and feel foreign the moment a customer switches to Hindi.
Inclusion, not automation
This isn't replacing a call centre. These businesses never had one. It's the first time a two-person shop can answer every customer who calls.
Cross-sector by nature
One product already serving health, education, retail and logistics. That reads as broad impact rather than a single niche — which is exactly what the IndiaAI Mission is backing.
The phone rings, Swaram picks up
A customer dials the business number. The call arrives over a SIP trunk from an Indian telecom provider, terminates into a LiveKit room, and the agent joins it as a participant. Press call and watch the whole loop run — including the tool calls it makes mid-conversation.
Build on SIP, not SDKs
Exotel, Plivo, Twilio, Ozonetel and Knowlarity all speak SIP. Keep the provider as a database row, not a branch in the code — switching becomes config, not a rewrite.
Numbers take days, not minutes
Indian DIDs need KYC and provider review. Treat "number pending" as a real onboarding state and put the business live on the widget meanwhile.
Forwarding is the cheap wedge
Owners already have a number customers know. Conditional forward on busy or no-answer means Swaram catches exactly the missed calls — no new signboard.
The website that talks back
One script tag on the owner's site. The loader injects a Shadow DOM widget and opens WebRTC from the parent origin, so there's no cross-origin microphone permission to get wrong. Click the orange button on the demo site.
—
Why not a plain iframe
A cross-origin iframe can't touch the microphone unless the parent page adds allow="microphone" to the tag. Every customer has to edit HTML correctly, and any CMS that strips attributes breaks it silently.
The install, in full
<!-- paste once, before </body> --> <script src="https://cdn.swaram.in/w.js" data-agent="ag_7f3c91" data-lang="hi,en,or" async></script>
Live in minutes — genuinely, for this module. Don't let that promise leak onto the phone-number path, where KYC takes days.
The dependency everyone forgets: TURN
On Indian mobile networks and corporate wifi, a meaningful share of WebRTC sessions can't connect peer-to-peer and must relay through a TURN server. LiveKit Cloud includes it; self-hosting means running coturn and paying relay bandwidth. Budget for it before it surprises you in production.
The app, split in two
The owner app is the real product — it's where the business lives and where every escalation lands. The customer doesn't get an app; they get a link and a QR code, because nobody downloads an app to call a kirana store.
Why the customer-side super app is a trap
A customer will not download an app to call a kirana store — they'll dial, or WhatsApp. A two-sided directory needs enormous liquidity before either side turns up, and it would put Swaram against JustDial, Zomato, Swiggy, ONDC and WhatsApp at the same time as selling a voice agent. That's two hard companies stapled together.
What the owner app earns its place doing
| Handover receiver | The owner's phone rings when the agent escalates, with the transcript already on screen. Run the logistics call to see it fire. |
| Business knowledge | Photograph the rate card, paste a URL, or let the agent interview the owner. Nobody types a catalogue. |
| Lead & booking inbox | Every captured lead and booking in one list, with the call it came from. |
| Wallet & plan | Minutes remaining, the 80% nudge, one-tap upgrade over UPI autopay. |
| Login | Phone and OTP. Email-and-password is the wrong door for this audience. |
Every capability, playable
These are the features behind the claims in the brief. Each one is a working demo, not a screenshot.
Speaks your customer's language
The same opening line, for the same business, in every language Swaram answers in. Click any card to hear it delivered.
Code-mixing is the real test
Indians don't speak one language per sentence. A global English-first model breaks on this; an Indian-language model is built for it. Here's an actual customer sentence, tokenised by language:
Detects, doesn't ask
The agent picks up the language from the first few words. No "press 1 for Hindi" — that menu is the thing customers hang up on.
Switches mid-call
If the customer switches, the agent switches. Numbers, prices and names stay intelligible either way.
Narrowband is the hard case
Phone audio is 8 kHz and noisy. The widget will always sound better than the phone line — test on the phone line, not the laptop.
Answers from your own business information
The single biggest lever on answer quality — and the thing owners will never do if it involves a form. Click a source to ingest it.
Extracted knowledge
0 factsAsk it something
The agent only answers from what it actually knows, and it says so when it doesn't.
Sign up and go live in minutes
No sales call, no demo booking, no onboarding manager. Press start and watch the clock — this is the whole journey from download to first answered call.
A face for your business, online
The same agent, same knowledge, same tools — with a face on screen. Sold as a premium add-on, for reasons the cost bar below makes obvious.
Cost per minute, relative
Illustrative ratio, not a quote — but the shape is right.
Anam bills on credits with a small free allowance. Video minutes cost many times what voice minutes cost, which is why they can never sit inside a ₹999 bundle.
How to sell it
| Add-on only | Never bundled into a plan. Separate line item, separate meter. |
| Hard minute cap | Cut off at the cap rather than billing through it. Surprise bills kill trust in this segment. |
| Website first | Video makes sense on a webpage or a kiosk. It makes no sense on a phone call. |
| Good fit | Clinics, salons, real estate, showrooms — where a face builds trust before a visit. |
Hands the hard ones to your team
The agent knows what it can't do. When it escalates, it passes the whole conversation across so the customer never repeats themselves. Press play to walk the fallback chain — including what happens when the owner doesn't pick up.
What gets passed across
| Full transcript | Everything said, in the original language plus a translation. |
| Why it escalated | The specific thing the agent couldn't decide or wasn't allowed to. |
| Customer details | Number, name, and anything captured mid-call. |
| Suggested action | What the agent would have done if it had permission. |
Rules, not vibes
Each vertical declares what forces a handover — a clinic never gives medical advice, logistics never promises a priority slot, coaching never negotiates fees.
Angry is a trigger
Repeated frustration, a raised voice, or an explicit "let me talk to a person" escalates immediately. Arguing with an upset customer is the fastest way to lose them.
Warm, not cold
The agent tells the customer it's connecting them and why. Silence during a transfer reads as a dropped call.
Every conversation saved and searchable
Search across every call this business has taken — transcripts, outcomes, phone numbers. Type anything, or try one of the suggestions.
See your results at a glance
What the owner opens the app to check. Deliberately four numbers, not forty — this is a shopkeeper's dashboard, not an analyst's.
Calls by hour
Where calls end up
The funnel the owner actually cares about.
Most asked questions
Fix the top three on your website and the call volume drops.
What we deliberately don't show yet
| Call quality scoring | On the roadmap. Don't claim it. |
| Conversion attribution | Needs the owner's sales data, which they don't have digitally. |
| Sentiment charts | Impressive in a demo, useless to a kirana owner. Cut it. |
Start free, upgrade in one tap
Drag the slider to burn through the free minutes and watch what the owner sees. The nudge fires at 80%, not at zero — running out mid-call is how you lose a customer.
At your current pace you'll run out in about two days. Upgrade now and nothing gets interrupted.
Why bundles, not a meter
A per-minute meter makes a shopkeeper anxious about every call — the exact opposite of what we want. A bundle is a known number they've already paid, so they stop counting and let the agent work.
Payment realities in this segment
| UPI autopay | Essential. Card-on-file recurring fails constantly here. |
| Wallet top-up | For owners who'd rather pay as they go than commit monthly. |
| Never hard-stop | Going over should degrade, not disconnect. Cutting a live call to enforce a limit is unforgivable. |
| Rupees, always | Dollar pricing reads as "not for me" instantly. |
Your business, your data
A voice agent hears things — names, phone numbers, health complaints, addresses. Being specific about what happens to that is not optional, and under India's DPDP Act it's law.
What we store
| Transcripts | Yes — that's the searchable record the owner is paying for. |
| Audio recordings | Optional, off by default, retention set by the owner. |
| Customer numbers | Yes, they're the lead. Visible only to that business. |
| Business knowledge | Yes, scoped to that business alone. Never pooled across customers. |
What we commit to
| India residency | Data stays in Indian regions. Also a latency win, not just compliance. |
| No cross-training | One business's conversations never train anything another business touches. |
| Consent on record | Calls announce they're recorded, in the customer's language, before anything is captured. |
| Deletion means deletion | Owner deletes, it's gone — including from backups within the stated window. |
Three doors, one room
This is the decision that keeps Swaram buildable. LiveKit has a SIP bridge, so a phone call terminates into a room as just another participant — identical to a browser visitor or an app user. One agent worker serves all three. Build it any other way and the STT→LLM→TTS loop gets written three times and drifts three ways.
The pipeline, one turn at a time
Press play to walk a single customer utterance through the runtime. The budget is under one second from the moment they stop talking to the moment they hear a reply.
A vertical is four things, not a fork
| Layer | What changes per industry | Current pack |
|---|
Dependencies — click one to break it
Every box below is something outside our control. Toggle any of them to failing and watch which modules go down, degrade, or survive. This is the map worth arguing about before writing code, not after an outage.
Owner-side dependencies
| Login | Phone and OTP. Emails get forgotten and passwords reset weekly in this segment. |
| Business knowledge | Photo of the rate card, a URL, or a five-minute onboarding call. Never a form. |
| Number KYC | Blocks phone go-live for days. Ship them live on the widget meanwhile. |
| Tool / API access | They don't have one. Swaram owns the data. |
| Handover reachability | App push → owner mobile → WhatsApp summary. Never a dead end. |
| Payment mandate | UPI autopay, or ₹999 recurring silently fails. |
Design rules that fall out of this map
| Two of everything | Two telephony providers, a fallback model, a fallback TTS. Behind one interface each. |
| Degrade, don't die | Tools unavailable? The agent still answers and takes a message — a far better call than a busy tone. |
| Video is an add-on | Avatar minutes cost many times voice minutes. Never inside ₹999, and hard-capped. |
| Colocate in India | Every network boundary is latency you can't get back. |
The number to settle before launch
Starter is ₹999 for 1,500 minutes — ₹0.67 a minute, covering telephony, speech, the model, media relay, infrastructure and payment fees before a single rupee of margin. Drag in real vendor quotes and see whether the plan survives.
Plan
Cost per minute
—
Per business, per month
| Revenue | — |
| Minutes actually used | — |
| Variable cost | — |
| Payment fees | — |
| Gross profit | — |
| Gross margin | — |