The answer lives in your systems
Prices, stock, order state, appointment slots, account balances. The agent must read the source of truth on every turn, not a copy that went stale last week.
Custom agents and projects
A complete phone agent, wired into your systems, delivered working.
Speech is the easy part. The project is everything underneath it: your catalog, your CRM, your calendars, your policies, and the bridge that lets the agent read and write them in milliseconds while a customer is on the line. We design, build, seed and certify the whole stack, end to end.
- Inbound and outbound, on your number or a new one
- Reads and writes your real systems, no copied data
- Hebrew, Arabic, Russian and English, natively
- Monitored, measured and maintained after go-live
- 1Discovery
- 2Architecture
- 3Systems map
- 4Integrations
- 5Agent design
- 6Knowledge
- 7Certification
- 8Operation
When a custom agent is the right answer
The self-serve product covers a receptionist, bookings and order status out of the box. A custom project is for the cases where the agent has to finish a real process inside your own systems, under your own rules.
A call is only done when the process is done
Book, change, cancel, verify, update an address, open a ticket, send a payment link, write the outcome back. We measure how many processes ended without a human, not how many calls were picked up.
Rules, authority and compliance are yours
What the agent may decide alone, what it must confirm, and what always goes to a person with full context. Recording, transcripts, audit and retention follow your regulation.
The stack, top to bottom
This is what runs under one call. The two layers at the top are the same for every project. The core, the modules and the connection to your systems are what we design and build for you.
Edge
Edge
Telephony
- Your existing number or a dedicated one
- Answers immediately, no queue
- Parallel calls, no busy line
- Inbound and outbound, call recording
- Serves customers who only have a phone
Voice engine
- Speech-to-speech model: audio in, audio out
- Natural Hebrew, Arabic, Russian and English
- Detects the speaker's gender and adapts the address
- End-of-speech detection and interruption handling
Core
Core
Your agent
- Goal, persona and playbook written for your process
- Advises, sells, books, verifies and resolves
- Knows your business, branches and policies
- Honest when it does not know: checks or hands off
Modules
Modules
CRM and caller memory
Builds a relationship
- Recognizes the caller by number, cached and pre-fetched
- Extracts name and email while the person talks
- The summary is loaded on the next call
Catalog and inventory
What is on your site
- Every category with specs and price
- Promotion and availability in real time
- Compares and narrows down between models
Knowledge base
The agent's brain
- Hybrid search: semantic plus keyword
- Branches, hours, terms, delivery, warranty
- Retrieved inside the call, without stopping
Channels
During the call
- Sends a link to complete the order or pay
- Receives a link or photo from the caller
- Summary and documents by WhatsApp or email
Your systems
Your systems
Sync from your systems
One source of truth
- ERP, CRM, e-commerce platform, calendars, ticketing
- Catalog, prices, promotions and stock
- Reloaded on every call, without a restart
- Read and write, with a separate permission per action
Handoff to a person
When a person is needed
- Your call center, queue or the nearest branch
- Transfers with the full context of the call
- A planned outcome, not a failure
What makes it answer in milliseconds
A phone conversation forgives nothing. Every lookup the agent makes sits inside the pause between the caller's last word and the agent's first. We engineer each integration so that pause stays natural.
- 01
Fast tools skip the model turn
A known lookup is a deterministic call to a local tool or a signed bridge, not a question to another AI. No extra reasoning, no second round trip.
- 02
Data lives next to the call
Catalog and knowledge are shipped to the gateway host as indexed local databases. A category or branch lookup returns in single-digit milliseconds; nothing crosses the internet mid-sentence.
- 03
Work starts while the phone is still ringing
The caller lookup fires during signaling, and the voice session is pre-warmed in the ring window so the greeting is ready on pickup.
- 04
Lookups begin before the sentence ends
When the partial transcript already shows a branch or product question, the lookup starts early and the answer is ready when the caller stops.
- 05
Warm indexes and short caches
The embedding model is loaded once at boot and kept warm. Availability and caller data are cached for seconds, so repeated questions cost nothing.
- 06
Acknowledgement in the agent's own voice
The instant a slower tool starts, a short acknowledgement plays in the same voice, so the caller never hears dead air. Every source has a deadline; if it is late the agent says so instead of freezing.
We do not guess at latency. Every certification call is recorded in stereo and decomposed per turn: end-of-speech detection, tool time and generation time, so we know exactly which part of each integration we tuned.
How a project runs
Eight phases, each with something you can hold at the end. The first seven take a project from a call map to your live line, usually through a pilot number within weeks. The eighth is how the agent is run from then on.
- 01
Discovery
Define the goal and what "done" means
We map the calls you actually receive, pick the processes worth automating first, and agree on the split of authority: what the agent decides alone, what it confirms, what always goes to a person.
- Which calls, at what volume, and what a resolved call looks like
- What the agent may promise and what it must never promise
- Which languages, hours and channels the agent covers
- 02
Architecture
Design the stack for your case
We decide what is an always-on fact, what is retrieved from the knowledge base, and what is a live lookup in your system. Each choice has a cost in tokens and milliseconds, so we set a prompt budget and a latency budget per turn before writing a line.
- Prompt budget: what the agent always knows versus what it looks up
- Which lookups must be local, which can be a bridge call
- Where credentials live, and what runs on your hardware
- 03
Systems map
Map the internal systems and data sources
ERP, CRM, e-commerce platform, calendar, ticketing, payment provider. For each one: is it the source of truth, is it read or write, what is the API, who owns access, and how fresh does the data have to be.
- First-party sources only; no directories, no scraped copies
- Read paths that can be made fast, write paths that must be confirmed
- IP allowlists, tokens and vaults arranged with your IT
- 04
Integrations
Build the integration and the voice bridge for each one
Every system gets its own bridge: a small, deterministic layer between the agent's tools and your API. Reads are cached, indexed and pre-fetched so they answer in milliseconds. Writes are confirmed, idempotent and safe to replay.
- Catalog sync into a local index, refreshed on your schedule
- Caller lookup and pre-warm during the ring window
- Each tool proven directly, without voice, before it touches a call
- 05
Agent design
Write the goal, persona and playbook
Name, voice, tone and register. What it says first, how it handles a hesitation, an objection, a correction, an angry caller. How it announces a lookup, when it asks one question instead of three, and when it stops and hands off.
- One question per turn, short sentences, never repeats itself
- Gender-aware address in Hebrew from the first sentence
- An honesty ladder: answer, check, or hand off, never invent
- 06
Knowledge
Seed the knowledge base and the catalog
Branches, hours, policies, delivery, warranty and product knowledge are written in the caller's vocabulary, indexed for hybrid search, and sized to the prompt budget. The catalog is synced from your platform with its real categories, attributes and prices.
- Every fact traced to your site, API, export, or you
- Retrieval probes: the right branch, model and policy come back first
- A refresh path that runs without a restart
- 07
Certification
Certify on real calls, pilot, then roll out
We run scripted and unscripted calls, review each one from the stereo recording with a per-turn latency decomposition, and fix at the layer that failed: data, bridge or playbook. Then a pilot on a dedicated number, then your line.
- Per-turn latency: detection, tool and generation, measured not estimated
- Content checks: right facts, right tone, right handoff
- Resolution rate on the pilot before the switch
- 08
Operation
Run, monitor and maintain
After the switch, the agent lives inside a loop. Every call is recorded, logged and written back. Watchdogs protect each live call. Sampled calls are reviewed from the recording. Content, integrations and the gateway are updated on a schedule, each through a canary and with a rollback. You run it with our tooling, or we run it for you.
- Call log with duration, turns, tool calls, tokens and cost per call
- Regular quality review from recordings, fixed at the failing layer
- Content and gateway updates ship through canary, drain and rollback
After go-live: how the agent is run
Going live is the middle of the project, not the end. Three things happen from then on: every call is watched, every live call is protected, and everything the agent depends on is maintained on a schedule.
Every call
Monitored
- Recorded in stereo, transcribed and summarized
- Logged with duration, turns, tool calls, tokens and cost
- Outcome written back: a CRM note, a ticket, a webhook or an email to your team
- Sampled calls reviewed from the recording with a per-turn latency breakdown
- Usage and cost tracked per call, with caps
During the call
Protected
- Dead-air watchdog: checks in, then ends the call gracefully instead of dropping it
- Stalled-model watchdog: nudges the session or recovers it
- Failure breaker: a source that keeps failing is not retried every turn; the agent says so and moves on
- Reconnect recovery when the voice session or the media stream drops
- One structured anomaly stream, so any defect on a call is found with a single search
On a schedule
Maintained
- Knowledge base and catalog re-synced from your sources, with a diff reviewed before it is pushed
- Playbook and persona changes are versioned and regenerated, never edited by hand on the box
- Gateway updates go to one canary agent first, then the rest; drain mode lets active calls finish before a restart
- Every update carries a rollback and reports the version that is really running
- Prompt size measured as rendered after every content change, so nothing silently falls out of budget
The operating rhythm
After every change
A certification call, reviewed from the recording before the change is called done.
Every week
A call-quality review: a ranked list of defects with evidence, each fixed at the layer that failed, whether data, bridge or playbook.
Every month
Resolution rate, handoff reasons and cost per resolved call, plus a content refresh from your sources.
You can run all of this yourselves with the control panel, the transcripts and the results API, or we run it as a managed service. Either way it is the same dashboards, the same reports and the same rollback.
What you hold at the end
A running agent, and everything needed to keep it running without us in the loop.
A live agent on your number, inbound and outbound
Working integrations with a test battery for each system
A knowledge base and catalog with a documented refresh path
The persona, playbook and greeting, versioned
Recording, transcript and a summary of every call
A control panel with the call log, usage and cost per call
Handoff to your people with the full context
Alerts when a call needs a person's attention
An operations plan: review cadence, refresh schedule, update and rollback path
An operations handbook your team can run from
Projects like yours
Three shapes we have built. The names are left out; the stack is real.
Retail chain
Sales advisor for a national electronics retailer
Advises and sells across tens of thousands of products, compares models, quotes price, promotion and stock, knows every branch and its hours, and sends the order link by WhatsApp while the customer is still on the line.
- Catalog sync from the store platform
- Local product and branch index
- CRM caller memory
- WhatsApp during the call
- Handoff to the call center
Appointments
Receptionist that books straight into the calendar
Finds the right staff member and service, checks live availability, books, changes and cancels, and never offers a slot that is not really free.
- Bookings bridge to the calendar platform
- Staff and service catalog
- Availability cache
- Confirmation by SMS or WhatsApp
E-commerce
Order status and returns for an online store
Recognizes the customer, reads the order and shipment state, explains the next step, opens a return, and writes the outcome back to the store.
- Store orders bridge
- Caller verification
- Ticketing write-back
- Secure payment link when needed
Questions we get before a project
What is the difference between this and the self-serve product?
The self-serve product gives you a receptionist, bookings and order status with a few clicks and a playbook you edit yourself. A custom project adds integrations that do not exist yet, a knowledge base and catalog seeded from your systems, a persona and playbook written for your process, and certification on real calls. The same gateway runs both.
Which systems can the agent connect to?
Anything with an API or an export: ERP, CRM, e-commerce platforms such as Shopify, Wix and Magento, calendars, ticketing, payment providers, and internal databases. Each one gets its own bridge with the read and write scope you approve.
Does our data leave our systems?
The agent reads and writes in your systems; it does not copy your database. Indexes needed for speed, such as the catalog and knowledge base, live on the gateway host, which can be a machine you own. Recording, retention and access follow your policy.
How fast does it answer?
Local lookups return in milliseconds, and lookups that need your system start before the caller finishes speaking. We do not quote a single number: every certification call is recorded and decomposed per turn, and we tune the integration that is slow rather than the whole agent.
What happens when the agent does not know?
It follows an honesty ladder: answer from a verified source, check a source, or hand off to a person with the full context. It is never allowed to invent a branch, a price or a policy, and the certification calls check for exactly that.
How long does a project take?
Discovery and architecture take days. Integrations depend on how many systems and how their APIs behave. Most projects reach a pilot on a dedicated number within a few weeks and move to your line once the pilot numbers are in.
How do you know the agent is behaving well?
Every call is recorded in stereo, transcribed and logged with its tool calls, tokens and cost. We review sampled calls from the recording, not from the transcript alone: the transcript shows what was said, the recording shows how it sounded and how long each pause really was. Findings go into a ranked list with evidence, and each one is fixed at the layer that failed.
What happens when our data changes?
The knowledge base and catalog are re-synced from your sources on a schedule you set, and the diff is reviewed before it is pushed. Prices, stock and availability that come from live lookups need no refresh at all. Playbook and persona changes are versioned and regenerated, so the live agent always matches a known version.
Who maintains it afterwards?
You can. Every project ends with a documented refresh path for the knowledge base and catalog, a versioned playbook, a control panel and an operations handbook. We also run agents as a managed service: weekly quality reviews, monthly resolution and cost reports, and updates shipped through a canary with a rollback.
Tell us about the process you want finished on the phone
Bring the calls, the systems and the rules. We come back with a call map, an architecture and a pilot plan.