A conversation runtime, not a bot builder.
Most voice vendors sell you a flowchart editor and hope. Nivākya runs a governed runtime: every call is admitted, scheduled, policy-checked, measured and written to an audit spine you can query afterwards. You get four surfaces on one model — operations, design, simulation and analytics — and nothing is retyped between them.
Every call is three moves. We publish the budget for all three.
A runtime is only as good as the worst millisecond inside it. Nivākya splits the turn into hear → decide → act, gives each stage its own latency budget, and alarms on the stage — not on the average — when the budget is blown.
Streaming ASR with true barge-in. The agent yields the floor 40 ms after the caller starts a word, so interruptions survive instead of being talked over.
Intent, entity and policy resolution against your price list, calendars, DNC register and knowledge base. Retrieval is pre-warmed per active campaign, so nothing cold-starts mid-sentence.
Speech synthesis, tool dispatch and CRM write issued in parallel. The deal record usually lands 300 ms before the caller finishes saying goodbye.
Measured at the carrier edge, caller-to-agent, across 4.2M production calls in the trailing 90 days. Regional splits sit on the status page.
Six layers, each one independently observable.
Nothing in Nivākya is a monolith you have to trust. Every layer emits metrics, has its own failure mode, and can be replaced — including the speech models, the LLM routing and the telephony carrier underneath.
Carrier-grade ingress
SIP trunking, PSTN failover and number portability across Mumbai, Frankfurt, Ashburn and Singapore edges. Calls are admitted, tagged with consent state, and pinned to a region before a single model sees audio.
Hear and speak, swappable
Streaming ASR, diarisation, endpointing and neural TTS, with per-language voice routing. Bring your own model or use ours; the runtime treats either as a pluggable provider with the same telemetry.
Rules that cannot be talked around
Every proposed response passes a deterministic guardrail pass before synthesis. Disclosure requirements, price floors, DNC lists and escalation triggers are compiled to fast predicates — evaluated in 9 ms, not prompt-engineered.
Actions with a paper trail
Typed tool calls to your CRM, calendar, ERP and payment rails, with schema validation, idempotency keys and a per-tool retry ladder. A failed booking retries three times and then hands off — it never silently disappears.
Context that survives the hangup
Caller profiles, open tickets, prior objections and grant-state are resolved before the first word and cached in a per-region store. Memory is versioned alongside the agent, so a rollback restores the facts the old version expected.
Append-only, queryable, hash-chained
Every turn, tool call and policy decision is appended with a SHA-256 chain and shipped to your own storage. Compliance queries 90 days of conversations in console time; we hold the chain, you hold the keys.
Rendered from the same trace object the console, the rehearsal report and the warehouse export all read from. There is no second copy of the truth.
An agent is an asset with a change history.
Prompt, tools, guardrails, voice, memory schema and escalation policy travel together in one immutable version. Nothing reaches production without a named reviewer, a passing rehearsal suite and a rollback path that has been tested at least once.
Draft
An editor branches from production and changes anything — a sentence of persona, a price floor, a calendar route. Every edit is captured with the author and the timestamp, never overwritten.
Review
A second person reviews a semantic diff: which behaviours changed, which guardrails moved, which tools were added. Sign-off is a named approval, not a thumbs-up in a chat thread.
Canary · 5%
The version takes one in twenty live calls for 45 minutes. Task completion, escalation rate and policy denials are compared against the incumbent with automatic halt thresholds.
Full rollout
Traffic moves in 25% steps with a 15-minute soak between each. Any halt condition returns traffic to the incumbent without dropping a live call.
Rollback
One click, under 20 seconds, mid-call. The runtime re-reads the prior version's memory schema on the fly, so a caller never notices the change of brain.
A failing gate does not block the edit — it blocks the deploy. Engineers keep shipping drafts; only the runtime is opinionated about which one customers meet.
Read the governance modelSame runtime, four ways to run it.
Regulated buyers rarely want the same thing twice. Pick the floor you are comfortable with — the API, the agent semantics and the audit format do not change when you move.
| Model | Where it runs | Data residency | Typical p50 turn | Who it suits | Ops burden |
|---|---|---|---|---|---|
| Managed cloud Shared multi-tenant |
Nivākya control plane, 4 regions | Region-pinned on request | 0.60s | Teams shipping a first production agent in under two weeks | None — we page ourselves |
| Dedicated VPC Single-tenant |
Your AWS, GCP or Azure account | Your cloud footprint | 0.64s | Enterprises with a network team and a security review it must pass | Terraform module, we own upgrades |
| In-country residency Sovereign region |
India (Mumbai), EU (Frankfurt), UAE (Dubai), UK (London) | Audio, transcripts and derivatives never leave the border | 0.61 – 0.71s | Banking, insurance, health and public sector under DPDP or GDPR scrutiny | We operate, you audit |
| On-prem enclave Air-gap capable |
Your data centre, GPU nodes you own | No egress at all | 0.78 – 1.10s | Defence, payments infrastructure, telcos with strict interconnect rules | Joint runbook, quarterly model refresh |
Latency ranges measured on comparable 8-vCPU / 2-GPU node classes. Moving between managed cloud and dedicated VPC is a configuration change, not a re-implementation — the agent version hash stays identical.
Buy, build, or keep the queue.
We lose deals to in-house builds and we lose deals to incumbents. Here is where each option actually wins, written by the team that has to support the choice afterwards.
| Dimension | Nivākya | Build-your-own stack | Legacy CCaaS + IVR |
|---|---|---|---|
| Time to first live call | 9–14 days | 4–9 months, plus hiring | 6–10 weeks of flow configuration |
| Latency ownership | Published per stage, alerted on | Yours to discover, usually in production | Menu trees, not conversational turn-taking |
| What you can prove after a call | Hash-chained transcript, policy decisions, tool calls, redaction receipt | Whatever you remembered to log | CDR fields and a recording |
| Guardrails | Deterministic rule pass, 9 ms, mapped to policy clauses | Prompt instructions that can be argued out of | Script branching decided in advance |
| Pre-release testing | 10,000 simulated calls per publish, free | A QA team on a headset | Regression scripts replayed quarterly |
| Multilingual behaviour | Mid-call code-switching across 42 languages | One model, one locale, one integration each | Separate IVR per language |
| Where it wins | Boring, observable, defensible conversations at volume | Full control, full cost, full responsibility | Deep telephony estate and existing contracts |
| Cost shape | Per connected conversation + pass-through carrier minutes | Engineering headcount, forever | Per-seat, per-month, whether calls come or not |
The spreadsheet we hand to CTOs, including the parts where building is cheaper.
Open the model →What we were missing against LivePerson, and what shipped in response.
Read the analysis →Latency, completion and hallucination rates across six vendors on one test set.
See the numbers →Put the runtime on one real number.
Bring three recordings of calls you are proud of and three you are not. We will build the agent against them, rehearse it overnight, and show you the report before it rings anyone new.
Managed cloud. 1,000 free conversations in the first 30 days. No credit card, no ring light.