Practice 013 services
Agentic AI
Systems that plan, use tools and finish the job.
Agent runDone
- Plan3 steps
- Tool · crm.lookupok
- Tool · calendar.bookok
- Check · verify resultpass
AI-native engineering · United States · Middle East
Every kind of AI. Built, trained and run.
Agentic systems, custom LLMs, speech and vision models, inference engineering, enterprise deployment and the products on top. One engineering team, from the model weights to the production line.
Engineering in India. Deployed on your servers, in your cloud or in-country when the data cannot leave.
- Training · epoch running
- Agents · tool call verified
- Inference · p95 within budget
- Review queue · one for a person
Scroll
- Clients in the United States and the Middle East
- Our own models, inference stack and voice platform
- Products in production: RingTrunk and CallWhiz AI
- On-premise, air-gapped and sovereign deployments
- Agentic AI development
- AI workflow automation
- Conversational AI and copilots
- Custom LLM training and fine-tuning
- Custom ASR and TTS training
- Computer vision and multimodal AI
- Data, labelling and synthetic data
- Evals, red-teaming and guardrails
- Inference engineering and optimisation
- On-device and edge AI
- Enterprise and sovereign AI deployment
- AI product and platform development
- Voice agent development
- Knowledge systems, RAG and enterprise search
- Document AI and intelligent processing
- AI strategy and transformation
- Managed AI operations
- Arabic-first AI
- Answer engine optimisation
- Agentic systems
- Computer-use agents
- Reasoning models
- RL post-training
- Small models and distillation
- Inference optimisation
- Real-time speech
- Vision-language models
- On-device models
- Context engineering
- Evals and guardrails
- Synthetic data
- Sovereign AI
01 · Capabilities
Five practices. Nineteen services. One team.
Agents, models, inference and infrastructure, products, and the strategy and operations around them. Pick the job; we do it end to end.
Practice 025 services
Models
Language, speech and vision models trained on your data.
Training · epochEval on your data
LossYour recordings
Practice 033 services
Inference and infrastructure
Serving that is fast, cheap and inside your boundary.
Serving · p95 latencyOptimised
- Baseline
- Quantised
- + Speculative
- + Cached
Practice 044 services
Products
AI-native applications, shipped and maintained.
Product · v1Shipped
- Buildpassing
- Evalsgating the release
- Deployinside your boundary
- Runbookshanded over
Practice 054 services
Strategy and operations
Knowing what to build, then running it.
QueueA person holds it
- CallBooked
- InvoiceTotal flagged
- EmailRouted
02 · Who we work with
All eight, in their wordsIf you are one of these people, this site was written for you.
Eight people we meet on the first call, across every kind of AI work. Each page is in their words: what they are trying to get done, what worries them, and what the first month looks like.
- 01 · Operations
Head of Operations or COO
Claims, orders, onboarding, invoices, tickets: the same steps hundreds of times a week, and a headcount that must not grow with them.
Read the page - 02 · Customer experience
Head of Customer Experience or Contact-Centre Director
Callers in Gulf Arabic and English on the same call, an IVR they abandon, and agents on the same five requests all day.
Read the page - 03 · Product engineering
CTO, VP Engineering or Head of Product at a software company
A promised AI feature, a first version on a frontier API, and a bill and a p95 that scale faster than revenue.
Read the page - 04 · Regulated technology
CIO, CISO or Head of Technology in a regulated institution
A board that wants AI, a regulator that wants controls, and a data-residency rule that every vendor's proposal breaks.
Read the page - 05 · Models
Head of AI, Head of Data Science or ML Lead
A mandate that says 'our own models', a small team, a GPU budget, and data that needs a pipeline before it trains anything.
Read the page - 06 · Voice products
Founder of a voice or speech product
A demo investors like, and a dialect, a latency or a cost-per-minute problem between it and a business.
Read the page - 07 · Public sector
Director of Digital Transformation or AI Programme Lead in the public sector
A national strategy with AI targets, a programme year to show something live in Arabic, and data that may not leave the country.
Read the page - 08 · Transformation
Chief Transformation or Digital Officer, or Head of AI Strategy
A board mandate, a budget that is now core operating expense, pilots that stall, and business units buying different tools.
Read the page
01 · Operations
Head of Operations or COO
Claims, orders, onboarding, invoices, tickets: the same steps hundreds of times a week, and a headcount that must not grow with them.
Read the page03 · One stack, any job
Watch a job move through the stack.
A call, an invoice or a camera frame. Same layers, same people on the queue, same log at the end. Pick one.
Run log — An inbound call
- 01[ your data ]Audio arrives on your number over SIP.
- 02[ models ]Custom ASR transcribes Gulf Arabic as it is spoken, streaming.
- 03[ inference ]Served on your servers, quantised, at conversation latency.
- 04[ agents ]The agent finds a slot, books it and checks its own work.
- 05[ channels ]Confirmation by SMS. The CRM record is updated.
- 06[ your outcome ]Call handled and logged. Recording, transcript and score attached.
- Outcome: Handled by the agent · logged · nothing for a person
Run log — An invoice
- [ your data ] A PDF lands in the shared inbox.
- [ models ] Layout-aware extraction reads the fields, handwriting included.
- [ inference ] A small fine-tuned model, batched, inside your network.
- [ agents ] Totals checked against the purchase order. One field uncertain.
- [ channels ] Record created in your ERP with the source line highlighted.
- [ your outcome ] Filed. The uncertain field waits in the review queue for a person.
Outcome: Filed · one field to a person
Run log — A camera frame
- [ your data ] A frame from the yard camera. No internet on site.
- [ models ] The detector finds the pallet; a vision-language model reads the label.
- [ inference ] On the edge box by the gate, in real time.
- [ agents ] The agent matches the label against the manifest.
- [ channels ] Alert to the floor supervisor. Event logged with the frame.
- [ your outcome ] A wrong pallet stopped before it left. Nothing to re-ship.
Outcome: Mismatch caught before dispatch
04 · Voice AI
Voice is where we go deepest.
Speech recognition and speech synthesis trained on your recordings, voice agents that hold real conversations on the phone, and the low-latency streaming, deployment and optimisation that make them work on a real line. All of it, by one team.
The turn · streamingTarget: under a second
- 01
Telephony
SIP and RTP in, jitter buffer, echo cancellation
- 02
Streaming ASR
Partial transcripts as the audio arrives
- 03
Endpointing
Silence plus a complete-looking sentence
- 04
Reasoning and tools
First sentence out while the rest is written
- 05
Streaming TTS
First audio on the first sentence
- 06
Playback and barge-in
Stops within a beat when the caller speaks
Voice capabilities
01
Custom ASR training
Speech recognition trained on your audio, your dialect and your line quality.
Read02
Custom TTS training
A voice that sounds like your brand, in the languages your customers speak, streaming.
Read03
Voice agents
Inbound and outbound conversations that finish a task in your systems and hand off when they should.
Read05
Deployment and telephony
On your servers, in your cloud, at the edge, and on the line your callers use.
ReadCustom ASR and TTS, by market: the Middle East, the United States, India.
05 · Answer engine optimisation
People ask an engine and read one answer. Make it yours.
Answer engine optimisation is how a company gets named, quoted and linked when ChatGPT, Perplexity, Google AI Overviews, Copilot and the assistants inside your customers’ own tools answer a question in your market. We do it the way we do everything else: measure first, engineer the content and the machine-readable layer, then track every answer.
- 01
Answer audit
We ask the engines the questions your buyers ask, in your markets and languages, and record who they name, what they cite and what they get wrong about you.
- 02
Content that answers
Pages rewritten so a model can lift the answer: one question per page, the answer first, the evidence under it, your name and your products written the same way everywhere.
- 03
The machine-readable layer
Structured data, llms.txt, a Markdown mirror of every page, clean sitemaps and pages that render without a browser, so crawlers and agents read what you meant.
- 04
Measured every week
A suite of real questions run against the engines on a schedule. Mentions, citations and share of answer tracked like latency, with a person reviewing the misses.
Why us
We build the answer engines’ kind of systems for a living: retrieval, evals, agents. We know what a model can quote and what it skips.
In English and Arabic. The engines answer in the language of the question, so the content has to exist in both.
06 · What is current
What is actually paid forThe techniques that matter this year, and what we do with them.
The field moves every quarter. This is what we are building with now, in plain words.
01
Agentic systems
Models that plan, call tools and finish multi-step jobs.
OursAgents with MCP tool access, checkers and human hand-off, in production.
Service02
Computer-use agents
Agents that operate software through the screen.
OursFor legacy systems with no API, with strict permissions.
Service03
Reasoning models
Models that think before answering, at a cost.
OursFine-tuned on your problems, routed only where they pay off.
Service04
RL post-training
Reinforcement learning with verifiable rewards.
OursFormat, tool use and correctness trained in, not prompted.
Service05
Small models and distillation
One job, one small model, a fraction of the cost.
OursDistilled from large teachers, served on your hardware.
Service06
Inference optimisation
Quantisation, speculative decoding, caching, routing.
OursLatency and cost engineered against your own load tests.
Service07
Real-time speech
Speech-to-speech models that listen and talk at once.
OursVoice agents that do not talk over people, in dialects.
Service08
Vision-language models
Ask questions of images and video in plain language.
OursInspection, reading and video events, on the edge.
Service09
On-device models
Small models running on phones, laptops and NPUs, with no cloud in the loop.
OursDistilled, quantised and shipped inside your app, with a hybrid path for the hard cases.
Service10
Context engineering
Retrieval done properly beats bigger prompts.
OursStructure-aware retrieval with citations and permissions.
Service11
Evals and guardrails
Measuring quality before and after every change.
OursTask suites from your cases, red-teaming, drift alerts.
Service12
Synthetic data
Model-generated examples where real data is scarce.
OursGenerated, reviewed by people, never used for evaluation.
Service13
Sovereign AI
Models and data kept in-country, on your hardware.
OursAir-gapped and in-region deployments for regulated sectors.
Service14
Answer engines
Buyers ask ChatGPT or Perplexity and read one answer.
OursContent and a machine-readable layer built to be the cited answer, measured weekly.
Service07 · What AI-native services means
Not software. Not an agency. The work, done.
An AI-native services company does the work instead of selling you a tool to do it. Agents are the workforce. People are on oversight. You pay for the outcome, delivered into the system you already use.
01
Your team spends its day on calls that follow a script and documents that follow a template.
02
Software does not fix it. Somebody still has to do the work.
03
Agencies will do it, at the price of people, forever.
The term comes from Y Combinator’s Summer 2026 request for startups. Most companies in the category rent a model and wrap it. We train the models, engineer the inference, build the products and run the systems, wherever the data has to sit.
“We don’t sell you an AI tool. We build the agents, train the models, engineer the inference and run the systems, wherever your data has to sit.”
| Software | An agency | AI-native services | |
|---|---|---|---|
| Who does the work | You do, with a tool | Their people, by the hour | Agents do it. Our people check it. |
| What you pay for | Seats and licences | Headcount, forever | Work delivered, on milestones |
| What comes back | A login | Output, at the speed of people | The result, in your CRM or your folder |
| When something is unclear | A support ticket | A phone call and a delay | A person in the queue, before it reaches you |
| Where the data sits | Their cloud | Their office | Your servers, your cloud, or ours |
08 · What separates us
We train the models. We engineer the inference. We run the systems.
Most AI service companies call an API and mark it up. That works until the domain is narrow, the dialect is rare, the latency budget is tight or the data cannot leave the building. We train our own language, speech and vision models, engineer the serving layer ourselves, run our own voice platform and telephony, and deploy the whole stack wherever the rules require. That is why we can promise the outcome instead of the tool.
Own models
Language, speech and vision models trained per client, on weights you own.
Own inference stack
Quantised, batched, cached and routed serving on the hardware you have.
Own voice platform and telephony
CallWhiz AI and RingTrunk, built and operated by us.
Deploy anywhere
On-premise, air-gapped, private cloud or in-country. A normal deployment for us.
Why it matters
When a vendor rents every layer, every problem is someone else’s. When we own the layers, the problem is ours to fix, that week.
The stack, end to end
- yoursYour data and systemsCalls, documents, records, camerasRuns on your servers, your private cloud, or ours
- oursModelsLanguage, speech and vision, trained per client
- oursInferenceQuantised, batched, cached, routed
- oursAgents and appsTools, evals, people on the queue
- oursVoice and channelsCallWhiz AI, RingTrunk, web, chat
- yoursYour outcomesRecords filled, calls handled, decisions made
Every layer between your data and your outcomes is ours: the models, the inference, the agents and the channels. Only the deployment boundary moves.
Who owns which layer
| Who usually owns it | Zingaro | |
|---|---|---|
| The agents and applications | a wrapper around someone else’s API | built in-house, with evals and a person in the loop |
| The models | rented, general-purpose | trained per client: language, speech, vision |
| The inference | whatever the vendor charges | engineered for your latency and cost, on your hardware |
| The voice platform and lines | a SaaS subscription and a carrier | CallWhiz AI and RingTrunk, ours |
| The deployment | someone else’s cloud | on-prem, air-gapped, private cloud or in-country |
09 · Industries
All industriesBuilt for the work in your industry.
The grind looks different in a clinic, a bank and a warehouse. The method is the same: one painful job, a pilot on your data, then it goes live.
01
Healthcare
Front desks, prior authorisations, claims and patient follow-up, with the data kept inside.
- Appointment calls and reminders
- Prior authorisations
- Claims intake
02
Banking and insurance
Onboarding documents, claims, collections calls and compliance monitoring, on-prem by default.
- KYC and onboarding documents
- Claims intake and triage
- Collections and renewal calls
03
Real estate
Lead qualification, viewings, tenant calls and document processing for brokers and developers.
- Inbound lead qualification
- Viewing scheduling
- Tenant and owner calls
04
Logistics and supply chain
Shipment queries, proof of delivery, exception handling and document checks between systems.
- Shipment status calls
- Proof of delivery and customs documents
- Exception chasing
05
Retail and e-commerce
Order calls, returns, catalogue data and customer questions answered from your own content.
- Order and returns calls
- Product questions
- Supplier invoices
06
Legal
Intake, document review, drafting from precedents and client updates, with the sources shown.
- Client intake
- First-pass document review
- Drafting from precedents
07
Government and public sector
Citizen services by phone in local languages, forms processed, records kept, hosted in-country.
- Citizen helplines
- Applications and forms
- Records and archives
08
Hospitality and travel
Reservations, guest requests and reviews handled around the clock, in the guest’s language.
- Reservations by phone
- Guest requests
- Confirmations and reminders
10 · How we work
From one painful job to a system in production.
Four steps. The text stays put while the scenes scroll past, each one a real screen from the work.
Models and agents handle the volume. People handle the edges. You always know which did what.
- 01
Discover.
One or two weeks with your team. The jobs listed, sized and ranked. A number on the first one.
Discovery · week oneRanked- Inbound calls
- Invoice intake
- CRM hygiene
- Leave alone
Eval suite · release gatePass- Groundednesspass
- Tool safetypass
- Formatpass
- Injectionpass
- 02
Build the pilot.
Fixed scope, fixed fee, 4 to 6 weeks. Real output on your real data, measured against the number.
The agent rebooks the appointment, the appointment field is extracted and highlighted, the CRM is updated, and the call is marked as handled by the agent with no escalation. - 03
Ship it.
Into production, inside your boundary if the rules require, with evals gating every release.
Your networkAir-gapped- Agents
- Models
- Voice
No outbound connection
Serving · p95 latencyOptimised- Baseline
- Quantised
- + Speculative
- + Cached
- 04
Run it, or hand it over.
We operate it with people on the queue and a weekly report, or your team takes it with the runbooks.
Queue · exampleA person holds the exceptions- Call · Inbound · appointment changeAgentBooked · CRM updated
- Document · Invoice intake · PDFPersonFiled · total flagged for a person
- Call · Outbound · follow-upAgentReached · note added
- Process · CRM hygiene · nightly runAgentRun complete · log written
- Email · Shared inbox · billing threadAgentRouted to billing
Weekly report · Week 38Work delivered- ✓Inbound calls answered and logged
- ✓Appointments booked and confirmed
- ✓Records updated in your CRM
Exceptions- !Caller asked for a refund: sent to your team
- !Invoice total unclear: source line highlighted
What changed- →“Cancel booking” added to the script
- →Model retrained on last week’s recordings
How an engagement works
Scope first. Then a price for the work, not for seats.
Fixed-scope build
A pilot or a product, scoped, fixed fee, billed on milestones. Nothing upfront.
Retained AI team
Engineers and researchers on your roadmap, month to month, inside your tools.
Managed operations
We run the systems, priced on the volume we actually handle. It moves with your work, not your headcount.
No seat licences, no minimum users, no annual lock-in. We quote once we have seen the work, and every milestone is invoiced only when it is accepted.
11 · Principles
Five things we will not bend on.
Read them once. Then hold us to them, in the pilot and every week after it.
01
We do the work.
You get the outcome in your systems, not a login and a training session.
02
We own the layers.
Models, inference, platform, lines. When something breaks, it is ours to fix, that week.
03
People hold the edges.
Agents take the volume. Reviewers hold the queue. You always know which did what.
04
Nothing invented.
No borrowed logos, no made-up numbers. The figures you see come from your own pilot.
05
It runs where the rules say.
On your servers, in your cloud, in your country. A normal deployment for us.
12 · Deployment
Your data does not have to leave the building.
01
On your servers
Air-gapped where required.
We install on machines you own, inside your network. Where the rules demand it, the system runs with no outbound connection and updates are carried in by hand. Your team keeps the keys.
Data stays inside your network
02
Your private cloud
We deploy into your account.
We deploy into your own cloud account, in your region, under your access controls. The data stays in your account. We get the access you grant, and nothing more.
Data stays inside your account
03
Ours
Managed, fastest to start.
We run it on infrastructure we operate. The right choice when the data is allowed to leave and you want the pilot running this month.
Managed by us, on our terms
In banking, insurance and healthcare, the rules decide where the data sits. We built for that first.
How each option works13 · Products we built
Two products, both ours. Proof we ship.
We did not integrate someone else’s stack. We built and run ours, which is why we can promise the outcome instead of the tool.
RingTrunk console+ Add number
| Number | City | Status |
|---|---|---|
| +91 80 •••• •••• | Bengaluru | Active |
| +91 22 •••• •••• | Mumbai | Active |
| +91 11 •••• •••• | Delhi | Active |
- SIP trunkRegistered
- POST /v1/calls200
RingTrunk
Indian phone numbers and SIP trunks for developers building voice agents. A console and a REST API. You point your own stack at it.
ringtrunk.comCallWhiz AI
Agents
- Front deskLive
- Collections follow-upLive
- After-hours coverLive
Campaigns
- Appointment remindersRunning
- Lead follow-upPaused
- Renewal callsDraft
Review queueA person holds it
CallWhiz AI
Our voice agent platform: agents, campaigns, a developer API, and the dashboards to run them.
callwhiz.ai14 · Where we operate
Someone is awake wherever the work is.
Clients in the United States and the Middle East. Engineering, and the telephony we run, in India. The clocks are live.
EasternLive
United States
Clients. Sales and delivery in your hours.
GulfLive
Middle East
Clients. Our deepest speech work, in Najdi and Gulf Arabic.
ISTLive
India
Home. The engineering team, and the numbers and trunks we operate.
15 · Writing
Notes from the work, and from the build.
The blog is for the people who own an operation. The engineering blog is for the people who will run what we ship.
Blog
All pieces21 September 2026 · 4 min read
Where the stock voices still fail: wealthy markets, under-served languages
Text-to-speech is excellent in about twenty languages and thin everywhere else. Some of the places where it is thin are rich, which is why custom voice training is a business. A map, with the caveats.
8 September 2026 · 4 min read
Where to start with AI when everything looks like a candidate
Most companies do not have an AI problem. They have a choosing problem. A short method for picking the first job, and the three tests it has to pass.
Engineering blog
All pieces12 September 2026 · 5 min read
Speculative decoding in production: when it pays and when it does not
Draft models and n-gram lookahead can cut per-token latency a great deal, or do nothing at all while burning compute. The mechanics, the workloads, and the harness we run before switching it on.
14 August 2026 · 6 min read
A latency budget for voice agents, millisecond by millisecond
A phone call forgives roughly a second of silence before it starts to feel wrong. Here is where those milliseconds go, and how we spend them.
16 · Questions
What people ask on the first call.
An AI-native services company does the work rather than selling software to do it. Models and agents are the workforce, people are on oversight, and the client pays for the outcome delivered into the systems it already uses. Zingaro AI is one, and it also trains the models and engineers the infrastructure the work runs on.
Zingaro AI takes on every kind of AI work that ships: agentic systems, workflow automation, assistants and copilots, custom LLM, speech and vision models, data and evaluation work, inference engineering, enterprise and sovereign deployment, AI products, voice agents, knowledge systems, document AI, answer engine optimisation, strategy and managed operations.
Zingaro AI uses both, chosen per job. Hosted frontier models are used where the data may leave and breadth matters; open-weight models fine-tuned and served by Zingaro AI are used where privacy, cost, latency or a narrow domain wins. Often a router decides per request, sending the common cases to a small model and the rest to a larger one or a person.
Zingaro AI deploys on the client's own servers, including air-gapped and in-country installations. That is a normal deployment for the company, not an exception, and it is built first for banking, insurance, healthcare and government.
Zingaro AI is accountable when a model or agent it runs gets something wrong. Uncertain work goes to a human queue before it reaches the client, every run is logged so it is clear which agent or person did what, and evaluation suites catch regressions before a release.
Zingaro AI does not publish prices because the cost depends on the work. The scope is set on a call and quoted after. Builds are milestone-based with nothing upfront, and managed operations are priced on the volume actually handled. There are no seat licences.
Answer engine optimisation is the work of making a company the answer that AI assistants and answer engines give: ChatGPT, Perplexity, Google AI Overviews, Copilot and the assistants inside customers’ own tools. Zingaro AI audits what the engines currently say, rewrites pages so a model can quote them, adds the machine-readable layer that crawlers and agents read, and measures the answers every week.
Book a call
Bring us the job you keep postponing.
Twenty minutes is enough to say whether we can take it.
- A pilot starts within 5 working days of agreed scope
- Nothing upfront
- No seat licences