Yatram

AI

AI that works on your fleet data — and can run on your own hardware.

Telematics produces a lot of text and a lot of numbers: alerts, tickets, trip histories, customer messages. AI is useful here when it saves an operator a click or drafts a reply they were going to type anyway. This page separates what Yatram does today from what we are still building — the roadmap items are marked, and nothing on the roadmap is running in the product yet.

Status: live

Available today

In the product now

Everything in this section is running in Yatram today and is covered by support.

Driver behaviour analytics

Overspeed, harsh acceleration, braking and cornering events are already captured per device and normalised per 100 km, so scores compare a long-haul driver fairly against a city courier.

Automated rule engine

Thirty-seven event types across speed, places, sensors, status and alarms, each able to notify by popup with sound, push, email, SMS, webhook — or send a command back to the vehicle automatically.

Knowledge base & canned replies

Support agents work from a searchable knowledge base and reusable replies, which is also the content an AI assistant will draw on.

Status: not yet available

On our roadmap

Not available yet

These are the AI capabilities we are working on. They are not in the product, they are not included in any current plan, and dates are not promised — each card shows where the work actually stands.

In development

AI assistant & site chatbot

A chat assistant that answers from your own knowledge base, documentation and fleet data — on the public site and inside the platform.

In development

AI-drafted support replies

Suggested ticket responses an agent reviews and approves before anything is sent. Nothing goes to a customer without a human saying yes.

Planned

Hybrid human + AI live chat

The bot handles first contact and hands over to a human with the full transcript, so nobody repeats themselves.

Planned

Visual chatbot flow builder

Drag-and-drop nodes — messages, reply buttons, input capture, conditions, webhook calls, handoff and ticket creation — so your team builds guided flows without code.

Planned

Predictive maintenance & risk scoring

Models trained on your own history to flag vehicles and drivers that need attention before a breakdown or an incident.

Planned

Ask your fleet a question

Natural-language reporting — "show me last week's overspeed by driver" — returning a real report you can export or schedule.

Roadmap capability · in development

A flow builder your support team can actually use

Not available yet

The illustration below is a design mock of the flow canvas we are building. It is not a screenshot of a shipped feature and it is not interactive.

Support flow · vehicle enquiriesDraftPreviewPublish
MESSAGEHi — how can we help?BUTTONSTrack my vehicleReport a problemCAPTURE INPUTVehicle numberCONDITIONVehicle found?HANDOFFTo a human agentCREATE TICKETSupport deskyesno
  • Built by the people who answer the messages. A support lead assembles a flow from message, button, input, condition and handoff blocks — no code, no regular expressions, no model prompt engineering required.
  • Draft first, publish when ready. A flow stays a draft while you test it, and only the published version is served to customers. Older versions remain available to roll back to.
  • A human is never more than one step away. Any block can hand the conversation to an agent or open a support ticket, carrying the full transcript and the vehicle or customer record with it.

Deployment options

Your data, your choice

Where inference runs is a decision for you, not for us. We intend to support both of the arrangements below and let each tenant pick. These are options we are building, not certifications or compliance guarantees — if you need a specific standard evidenced, ask us and we will tell you plainly what we can and cannot document.

Option A

Hosted model providers

Inference runs through a commercial model API. Nothing to install and no GPU to buy — the tenant administrator chooses the provider and supplies the key, and requests are made per conversation rather than by uploading your database.

  • Fastest to switch on: configuration only, no infrastructure
  • Per-tenant provider and key, set by the tenant administrator
  • Costs follow usage rather than a fixed hardware outlay
Option B

Self-hosted GPU inference

The model runs on a GPU server you own or rent, inside your own network. Prompts, telemetry and conversation history stay on your infrastructure, which is the deciding factor for operators under contractual or internal data-residency rules.

  • Tenant data used for inference does not leave your infrastructure
  • Per-tenant isolation: each tenant’s data and conversations stay separate
  • You control the hardware, the retention period and the upgrade schedule

Either way, tenant separation is how the platform is already built: each tenant's devices, customers and conversations are isolated, and that isolation is what an AI feature inherits rather than something bolted on beside it.

Want to be part of the build?

If a guided support flow or self-hosted inference would change how you run your fleet operation, tell us what it needs to do. We would rather design against a real requirement than guess.