Skip to content
Smriti AI

Smriti AI Labs

We build the systems that make AI useful.

We build scalable enterprise AI products and agentic platforms for businesses in India using Google Cloud, Gemini and Amazon Bedrock. Alongside our own voice models and AI harnesses, our engineering spans NVIDIA Hyperion-based ADAS and NVIDIA GPU-based inference.

Explore Enterprise AISee our work
Illustrative 3D render of a data-centre GPU module in exploded view: the module board with the GPU package, its memory and power delivery, and the heatsink lifted above it.
Illustrative 3D · GPU module

How the pieces fit

From a model to a working system.

1The model

A model is a capability.

On its own it can answer, summarise and reason. It does not know your business, your systems or when to stop.

2The harness

The harness makes it dependable.

Context, retrieval, tools, rules and evaluation decide what the model sees, what it may do and how its work is checked.

3The product

The product puts it to work.

An interface people use, knowledge you have approved and a person approving where it matters. That is when AI becomes a business system.

4The infrastructure

Engineered to run.

The same engineering carries into model serving on NVIDIA GPUs and into ADAS built on NVIDIA Hyperion.

Our businesses

Models, software and infrastructure, engineered together.

Smriti AI Labs builds scalable enterprise AI platforms, automotive intelligence, and accelerated inference systems—combining models, software, and infrastructure to engineer intelligence for the real world.

Intelligence for business

Current business

Enterprise AI Platforms

Scalable AI products and agentic platforms for businesses in India, built on Google Cloud and AWS, alongside our own voice models and AI harnesses.

Explore Enterprise AI

What we build

  • Agentic platforms and enterprise applications
  • Gemini-based systems on Google Cloud
  • Amazon Bedrock-based platforms on AWS
  • Smriti voice models and voice-enabled systems
  • AI harnesses, evaluation and deployment

Technologies we build with

  • Google Cloud
  • Gemini Enterprise
  • Gemini models
  • Agent Development Kit (ADK)
  • Cloud Text-to-Speech
  • AWS
  • Amazon Bedrock

Intelligence for mobility

Current business

Automotive Intelligence

Design, testing and refinement of ADAS solutions based on NVIDIA Hyperion and relevant NVIDIA DRIVE technologies, for Indian automotive applications.

  • ADAS solution design for intended operating conditions
  • AI software and model adaptation
  • Testing, evaluation and evidence-led refinement

Long-termResearch toward Level 5 autonomy. Long-term ambition — not a current capability.

Built with NVIDIA Hyperion · NVIDIA DRIVE

Explore Automotive Intelligence

Intelligence at scale

Current business

Accelerated Inference

Engineering of NVIDIA GPU-based inference systems: model deployment, serving and performance optimisation for AI applications.

  • Model deployment and serving on NVIDIA GPUs
  • Workload analysis and performance optimisation
  • Evaluation of response time, throughput, resource use and quality

Built with NVIDIA GPUs · NVIDIA inference software, selected per workload

Explore Accelerated Inference

Enterprise AI Platforms

Global AI foundations. Products built for Indian businesses.

We connect models, knowledge, voice, tools and application logic into complete business systems — on Google Cloud and AWS, alongside our own voice models and AI harnesses.

Explore Enterprise AI
  1. 01

    Business request

    “Which warranty claims need escalation this week?”

    Service operations team

  2. 02

    Approved knowledge

    • Warranty policy
    • Claims records
    • Service history

    Access checked per role

  3. 03

    Coordinated agents

    • Retrieve
    • Assess
    • Draft

    Orchestrated, with tool access

  4. 04

    Human approval

    Claims lead reviews proposed escalations

    Where the decision needs a person

  5. 05

    Business outcome

    • Escalations filed
    • Sources attached
    • Decision recorded

    Traceable end to end

DrawingFrom request to outcome
TypeIllustrative workflow
PracticeEnterprise AI

What a complete system connects

  • ModelsGemini models, foundation models available through Amazon Bedrock or our own — chosen per task.
  • KnowledgeAn organisation's approved documents and data, accessed with permissions.
  • VoiceSpeech in and out, with Smriti voice models or Google speech services.
  • ToolsThe business systems an agent may act in, through defined interfaces.
  • Application logicWorkflow rules, approval steps and the interface people use.
ResultA complete business systemDesigned, engineered and evaluated by Smriti.

Google Cloud

Google Cloud and Gemini

From business knowledge to coordinated action.

Gemini provides model capabilities. Smriti engineers how those capabilities become a product.

Google Cloud & Gemini practice

Google Cloud provides

Gemini Enterprise
Google's platform for organisations — the Gemini Enterprise app, connectors to company systems and prebuilt and custom agents.
Gemini models
Google's model family: the model capabilities a product draws on.
Agent Development Kit (ADK)
Google's open-source, model-agnostic framework for building, debugging and deploying agents.
Cloud Text-to-Speech
Google's managed speech-synthesis service.

BoundaryGemini Enterprise is not a model-training service, and not every Gemini model supports every tuning method. Model adaptation is chosen per task.

Smriti engineers

Product and interface
The experience people use, including review and approval steps.
Agent orchestration
How specialised agents divide work and hand off — built with ADK where it fits.
Integrations
Connections to business systems and approved knowledge, scoped to what each user may access.
Voice experiences
Speech built on Cloud Text-to-Speech or Smriti voice models, chosen per application.
Evaluation and deployment
Task-level checks before release, and the engineering to run the system.

AWS

AWS and Amazon Bedrock

Model choice becomes a business platform.

Amazon Bedrock is part of the foundation. The product experience and systems engineering are Smriti's work.

AWS & Amazon Bedrock practice

AWS provides

Foundation models available through Amazon Bedrock
A fully managed service giving access to foundation models from multiple AI companies.
Supporting AWS services, where relevant
Such as Amazon Bedrock Knowledge Bases, Amazon Bedrock Guardrails or Amazon Bedrock AgentCore — used only when a project requires them.

BoundaryAmazon Bedrock does not by itself provide every application requirement, permission boundary or scalability guarantee. That engineering is part of our work.

Smriti engineers

Business interface
The applications and workflows people actually use.
Agent or workflow logic
Orchestration, application rules, tool calls and hand-offs.
Model selection
Comparing candidate models against the task itself, not a generic benchmark.
Approved data and tools
Connections to the knowledge sources and systems the platform may use.
Evaluation and operations
Measurement, monitoring and improvement once the platform is in use.

Our own engineering

More than model integration.

Much of what makes an AI system useful is engineered around the model. We develop our own voice models and AI harnesses, refine models where a task requires it, evaluate against real tasks and build the applications people use.

Developed by Smriti

Voice models
Our own voice-model development, refined and evaluated for each application.
AI harnesses
Context handling, orchestration, tool integration, retrieval, workflow rules and deployment logic.
Model refinement
Adapting a model where the task requires it — and checking that it helped.
Evaluation
Task-specific test sets and checks, run before every significant change.
Application engineering
The interfaces, integrations and workflows people actually use.

Third-party models and services we integrate

Gemini models
Google
Foundation models available through Amazon Bedrock
AWS · multiple model providers
Cloud Text-to-Speech / Cloud Speech-to-Text
Google Cloud

Chosen per task and named as technologies we build with. They remain their providers’ products; Smriti does not claim them as its own.

Automotive Intelligence

ADAS designed for the environments it must face.

Smriti AI Labs designs, tests and refines ADAS solutions based on NVIDIA Hyperion for Indian automotive applications. We focus on how a system behaves within its intended operating conditions — and use evidence to guide refinement.

An abstract top-down drawing of a two-lane road with an unmarked edge. An ego vehicle with an indicative sensor field shares the road with a two-wheeler, a car and a three-wheeler, approaches a speed breaker and passes a vehicle parked partly in the lane, while a pedestrian stands at the road edge. Boxes mark road users a system would need to detect; the pedestrian is marked as needing attention.

  • Road user to detect
  • Needs attention
  • Sensor coverage, indicative
  • Planned path
DrawingMixed traffic, top-down
TypeIllustrative graphic
NoteNot test data

Design establishes the intent. Testing reveals the reality. Refinement connects the two.

Our current work, as a loop

  1. 01DesignFunctions and behaviour defined for the intended operating conditions.
  2. 02TestStructured testing of how the system actually behaves, including where it falls short.
  3. 03EvaluateEvidence on behaviour and limitations, documented against the design intent.
  4. 04RefineModels and software adapted from that evidence — then back to design.

NVIDIA Hyperion is NVIDIA’s compute and sensor reference architecture. We engineer ADAS solutions based on it — our contribution is the solution design, AI software, model adaptation, testing and evaluation. Using the platform does not by itself establish an automation level.

Discuss an automotive projectExplore Automotive Intelligence

Accelerated Inference

Intelligence needs more than a trained model.

Smriti engineers NVIDIA GPU-based inference systems for AI applications. We work on the execution layer: deploying models, understanding workload behaviour and refining how models run within the needs of the application.

  1. 01Entry point

    Application request

    A person or system calls the AI application, which has its own latency and quality needs.

  2. 02Smriti engineers

    Serving layer

    Routes each request and groups work into batches the hardware can run efficiently.

    • Routing
    • Batching
  3. 03Smriti tunes

    Optimised model execution

    The model runs on inference software selected for this workload, with the configuration and optimisation we engineer.

    • Software selected per workload
  4. 04NVIDIA provides

    NVIDIA GPUs

    The accelerated hardware the model executes on.

Evaluated across four dimensions

Measurements feed back into serving and execution choices

  • Response timeHow long a request takes, end to end.
  • ThroughputHow much work the system completes in a given time.
  • Resource useThe GPU memory and compute it takes to deliver that work.
  • Model qualityWhether outputs stay as good as the application needs.

No figures are shown here. Results are published only with the model, hardware, configuration, workload and measurement conditions. NVIDIA inference software — such as NVIDIA TensorRT, NVIDIA TensorRT LLM, NVIDIA Dynamo-Triton, NVIDIA Dynamo or NVIDIA NIM microservices — is selected per workload and named on a project only when it is used.

DrawingInference execution path
TypeSystem diagram
PracticeAccelerated Inference

Today, this is engineering work on NVIDIA GPU-based systems. Smriti-owned GPU cloud capacity is a future direction, not a current service. Our infrastructure direction

Discuss an inference projectExplore Accelerated Inference

Selected work

What we engineer, by project family.

Each entry is a capability overview: the kind of problem it addresses and what we engineer. They name no customers and claim no results — project write-ups are published only once they are approved.

Enterprise AI

  • Capability overview

    Gemini-powered enterprise agents

    Agents that work with an organisation's own knowledge and software, built on Google Cloud with Gemini models, Gemini Enterprise and the Agent Development Kit.

    Technologies: Google Cloud · Gemini models · Gemini Enterprise · Agent Development Kit (ADK)

  • Capability overview

    Amazon Bedrock-based business platforms

    Business platforms on AWS that combine foundation models available through Amazon Bedrock with an organisation's knowledge, tools and workflows.

    Technologies: AWS · Amazon Bedrock

  • Capability overview

    Smriti voice models and applications

    Voice-enabled systems that combine Smriti's own voice-model development with speech applications and, where relevant, third-party speech services.

    Technologies: Smriti voice models · Cloud Text-to-Speech · Cloud Speech-to-Text

  • Capability overview

    AI harnesses and evaluation systems

    The engineering around a model that makes it dependable: context handling, orchestration, tool integration, retrieval, workflow rules, evaluation and deployment logic.

    Technologies: Model-agnostic · Google Cloud · AWS

Automotive Intelligence

  • Capability overview

    NVIDIA Hyperion-based ADAS engineering

    Design, testing and refinement of ADAS solutions built around NVIDIA Hyperion and relevant NVIDIA DRIVE technologies, for Indian automotive applications.

    Technologies: NVIDIA Hyperion · NVIDIA DRIVE

Accelerated Inference

  • Capability overview

    NVIDIA GPU inference systems

    Model deployment, serving and performance engineering on NVIDIA GPU-based systems, shaped around what the application needs.

    Technologies: NVIDIA GPUs

View the portfolio

How we engineer

One way of working, from the problem to the system in use.

Every system we build goes through the same six steps, whichever business it belongs to. Model adaptation depends on the task: we tune or refine a model only where the task requires it — many problems are better solved through system design, retrieval, tools and evaluation.
  1. 01

    Understand the problem

    The users, the decision or task, the systems involved and what a good outcome looks like.

  2. 02

    Design the system

    Interfaces, workflow, model choices, data access and the points where a person stays in control.

  3. 03

    Develop and refine

    Application engineering and integrations. Models are adapted only when the task requires it.

  4. 04

    Evaluate

    Task-specific tests and measurements, including where the system falls short.

  5. 05

    Deploy

    Into the environment it must run in, with the monitoring needed to operate it.

  6. 06

    Improve

    Evidence from real use feeds the next iteration.

Improvement feeds back into understanding — the process is a loop, not a hand-off.

Careers and internships

Contribute to engineering work across AI and cloud computing.

Internships at Smriti are about contributing to our engineering work in AI and cloud computing. We do not list open positions here — read how internships work, then tell us what you would like to work on.

Internships at SmritiRegister your interest

Discuss a project

Bring us the problem. Let’s engineer the system.

Tell us what you are working on. Enquiries go straight to the founders, and we reply by email.

Useful to include

The problem
What you are trying to achieve, and for whom.
The systems involved
Applications, data, vehicles or infrastructure the work touches.
The constraints
Where it must run, timing, and the rules you work within.
Prefer a full page? Go to the contact page

Email us with a few clear paragraphs: the problem, the systems involved and the constraints. We reply to every message.

hello@smritiailabs.in