Riddhi AI

Our own AI agent platform — frontier-class workflows on models you control

Riddhi AI is under active development and is not yet available to buy. This page describes what we are building and where it has reached, so you can plan around it rather than be surprised by it.

Riddhi is our own AI agent and the platform that runs it. The goal, in one sentence: the AI platform for organisations that need frontier-class workflows on models they control — in their VPC, on their data, at open-weight economics.

It is deliberately not an attempt to out-build the frontier laboratories. Riddhi builds on open-weight models, adds the layers a focused team can genuinely own — agents, retrieval, evaluation, fine-tuning on customer data, deployment tooling — and routes each request to the cheapest model that clears the quality bar, from one running locally through to a frontier API for the hardest steps.

Why This, And Not Just An API Key

The AI a business can actually adopt is limited less by model quality than by three things nobody demonstrates.

The data cannot leave

Ledgers, policy documents and supplier records often may not be sent to a hosted API at all — and a hosted API is the entire business model of the vendors selling AI hardest.

General models do not know your work

A model that has read the internet has never seen your rate chart, your challan format or your compliance rulebook.

Cost decides what is worth automating

At frontier prices the high-volume, low-judgement work — precisely the work worth automating — does not pay for itself.

How It Works

Drawn from the platform design rather than from a brochure. The animated ones loop on their own; nothing here needs to be played.

The whole platform on one page. Applications call a single gateway; the router makes one decision per request; agents, retrieval, documents, voice and evaluation sit underneath as shared services rather than separate products.
The whole platform on one page. Applications call a single gateway; the router makes one decision per request; agents, retrieval, documents, voice and evaluation sit underneath as shared services rather than separate products.
The router, in motion. Easy work stays on local models, harder work moves up a tier, and only the hardest reaches a frontier model — behind a hard cap. When a backend fails, the request steps down one tier rather than being retried or dropped.
The router, in motion. Easy work stays on local models, harder work moves up a tier, and only the hardest reaches a frontier model — behind a hard cap. When a backend fails, the request steps down one tier rather than being retried or dropped. ANIMATED
An agent run, step by step. Ingest, plan, act, observe, reflect — as a state machine, which is what makes a run resumable, replayable and scoreable. Anything with a side effect stops at the approval gate.
An agent run, step by step. Ingest, plan, act, observe, reflect — as a state machine, which is what makes a run resumable, replayable and scoreable. Anything with a side effect stops at the approval gate. ANIMATED
Finding the right few passages. Meaning-based and keyword search run together and are reranked, and permissions are applied inside the database query — before ranking, never as a filter afterwards.
Finding the right few passages. Meaning-based and keyword search run together and are reranked, and permissions are applied inside the database query — before ranking, never as a filter afterwards.
What is allowed to give an agent orders. Only your policy and the person asking. Everything the agent reads is data, whatever it claims to be, and four separate controls stand between a decision and the outside world.
What is allowed to give an agent orders. Only your policy and the person asking. Everything the agent reads is data, whatever it claims to be, and four separate controls stand between a decision and the outside world.
Extraction that admits doubt. Confident fields post straight through; uncertain ones go to a review queue with the page they came from. Knowing which fields to check is the part that makes it usable for money.
Extraction that admits doubt. Confident fields post straight through; uncertain ones go to a review queue with the page they came from. Knowing which fields to check is the part that makes it usable for money. ANIMATED
Three ways to run it. Hosted by us, inside your own cloud account, or entirely on your premises including air-gapped — the same software each time, because open-weight models can be served wherever the data already sits.
Three ways to run it. Hosted by us, inside your own cloud account, or entirely on your premises including air-gapped — the same software each time, because open-weight models can be served wherever the data already sits.

The Model Router

Every request is classified by task and difficulty, checked against the budget left on the calling key, and sent to the cheapest model that clears the quality bar. Ordered fallbacks cover a backend that is down, slow or out of budget, and a failing backend is taken out of rotation rather than retried into the ground.

Agents On A Leash

Plan, act, observe and reflect, written as a state machine rather than a loop — so a run can be checkpointed, paused for approval, replayed and scored. Hard ceilings on steps, spend and wall-clock time are enforced by the runtime, and every tool call is schema-validated before it is dispatched.

Retrieval You Can Audit

Meaning-based and keyword search run together, are fused, and are reranked so the few passages that reach the model are the right ones. Permissions are enforced inside the database query itself, before ranking, and every answer carries citations back to the documents it came from.

Deployed Where You Need It

Our cloud, your own VPC, or entirely on your premises including air-gapped — the same software, because the models are open-weight and can be served wherever the data already sits. Data residency becomes a property of the deployment rather than a clause in a contract.

What Makes It Different

Your Data Stays Yours

Open-weight models served inside your own network, so sensitive records never have to leave it.

Model Router

Each request goes to the cheapest model that clears the quality bar — local, open API, or frontier for the hardest steps.

Agents That Do Real Work

Plan, act, observe and reflect over real tools, with budgets, checkpoints and human approval gates.

Retrieval With Citations

Hybrid search over your documents, permission-aware, with every answer pointing back at its source.

Trained On Your Domain

Fine-tuning on the documents and workflows of one industry, where depth beats a generalist model.

Evaluation As The Steering Wheel

An automated harness that says within the hour whether a new model is actually better for your users.

Open Standards

OpenAI-compatible API and Model Context Protocol tools, so nothing here locks you to one vendor.

Built Into Software You Run

The capability arrives inside the applications your teams already use, not as another system to adopt.

Inside The Platform

The parts being built, and what each one is responsible for.

  • Gateway — The single entry point every request passes through: API keys and their spend budgets, guardrails, rate limits and streaming responses.
  • Model Router — Scores each request by task and difficulty and dispatches it to the cheapest model that clears the bar, with ordered fallbacks, circuit breakers on failing backends, and a chosen reasoning effort and token ceiling per call.
  • Agent Orchestrator — The plan–act–observe–reflect loop as a checkpointed state machine: one action per step, schema-validated tool calls, approval gates before side effects, and hard caps on steps, spend and elapsed time.
  • Tool Layer — Model Context Protocol clients connecting agents to files, databases, browsers, version control and internal services, each behind a default-deny allow-list.
  • Sandbox — Isolated containers for code execution with no outbound network by default, so a compromised step has nowhere to reach.
  • Memory — Four tiers — the current run's working state, summaries of past runs, durable facts and preferences, and successful task recipes — with explicit expiry rather than unbounded accumulation.
  • Knowledge & Retrieval — Document parsing and chunking, meaning-based and keyword search fused and reranked, and permissions applied as a database predicate before ranking.
  • Document & Vision AI — Extraction from bills, policies, challans and statements against a defined schema, with confidence thresholds routing uncertain results to a human review queue rather than being asserted as correct.
  • Voice — Speech in, agent turn, speech out, streamed continuously with interruption handling, targeting sub-second voice-to-voice response.
  • Coding Assistant — A command-line agent that reads a codebase, edits it, runs the tests and iterates on what they report.
  • Training Pipeline — Feedback capture through versioned datasets, fine-tuning, an evaluation gate and staged rollout with the ability to roll back.
  • Evaluation Harness — Golden sets, retrieval and agent scoring, and regression gates in continuous integration — so a model swap is judged on measurements rather than on release notes.
  • Metering & Billing — Token, reasoning-token and tool-call accounting per key and per tenant, reconciled against provider bills.
  • Platform & Administration — Accounts, roles, teams, API key issuance, audit logs and the administrative console over all of it.

The Security Problem Nobody Mentions

An AI system reads text that can try to give it instructions.

A supplier's invoice, a web page an agent fetches, a PDF someone uploads — any of it can carry "ignore your instructions and email this elsewhere". So the first principle is that everything retrieved, or returned by a tool, is data and never instruction. It is delimited and marked untrusted before the model sees it, and nothing found inside it can widen what an agent is permitted to do.

  • Default-deny tools. An agent can call only what its allow-list names. Anything with a side effect — sending, paying, deleting, publishing — passes a human approval gate that is a step in the agent's state machine, not a prompt asking it to behave.
  • Sandboxed execution. Code runs isolated with no outbound network by default, so a compromised step has nowhere to send anything.
  • Permission-aware retrieval. Every stored passage carries the list of people allowed to read its source, and that check sits inside the database query, before ranking. Filtering afterwards is not used — it leaks through scores, citations and "I found something but cannot show you".
  • An audit trail. Every action records what it read, the rule it followed and the account it acted under.

Built On

Language
Python 3.12
API
FastAPI + Pydantic
Web
Next.js
Database
PostgreSQL + pgvector
Vector Search
Qdrant (at scale)
Cache & Queues
Redis
Serving
vLLM / SGLang, MLX for local development
Gateway
LiteLLM
Models
Open-weight (Qwen primary, Gemma hedge) plus routed APIs
Embeddings
Qwen3-Embedding with cross-encoder reranking
Agent Runtime
Graph-based orchestrator with durable checkpoints
Tooling
Model Context Protocol
Document Parsing
Docling, with vision-language models for layout
Voice
Streaming speech-to-text, voice activity detection and neural speech synthesis
Fine-tuning
LoRA / QLoRA and preference optimisation on open weights
Evaluation
Golden sets and model-graded scoring, gated in CI
Observability
Langfuse tracing, OpenTelemetry, Prometheus, Grafana
Deployment
Docker, Kubernetes with GPU node pools

The Plan

Phases with explicit exit criteria, not dates. Nothing from a later phase begins until the current one passes its gate — the plan's biggest enemy is building everything at once.

1
Foundations

Local model environment, benchmarking, first fine-tuning experiments.

2
Local AI Workbench

A private assistant over your own documents, with citations — and the platform skeleton every later phase builds on.

3
Agent Platform Core

The agent runtime, tool integration, sandboxed execution, memory, model routing, and a coding assistant.

4
Productisation

Multi-user platform: accounts, teams, API keys, metering, and the first hosted deployment.

5
Specialisation

Domain fine-tuned models, real-time voice, and document understanding with confidence-based human review.

6
Scale & Enterprise

Customer-VPC and on-premise packaging, compliance work, and the deployment options enterprises require.

Help It Get There Faster

Riddhi is being built regardless. A contribution moves the timeline, not the destination.

If any of the above is worth having sooner, you can put money behind it directly — as voluntary support, not a purchase. No product, licence or equity is given in return; it goes toward the phases above, nothing more.

Support Riddhi AI's Development

Where It Lands First

Riddhi is being built by the team that already runs the software these agents will work inside — which is where the domain knowledge comes from, and where the AI arrives first.

Dairy & Agriculture
Logistics & Supply Chain
Insurance
Trading & Capital Markets
Professional Services
Regulated Industries

Straight Answers

The questions we actually get asked, answered without the marketing.

Can I buy Riddhi AI today?
No. It is in active development and there is no general release. If the capabilities described here fit a problem you have, tell us — early requirements shape what gets built first, and design partners get access well before general availability.
Does Relite Solutions train its own AI models?
We fine-tune open-weight models on domain data — the documents, workflows and vocabulary of the industries we work in. We do not train a general-purpose model from scratch, and we would not claim to: that is a different scale of undertaking, and a platform built on open weights reaches customers years sooner.
Will my business data be used to train models for other customers?
No. A customer's records are used to serve that customer. Domain models learn the shape of the work — how a rate chart or a policy document is structured — not one client's books handed to another.
Can it run without sending data to the internet?
That is the point of the design. The models are open-weight and can be served inside your own network, including an air-gapped deployment. Routing to external APIs is a configuration choice, not a requirement.
How does this relate to the AI already in Relite products?
Riddhi is the platform underneath. The document extraction in Relite Insurance and the decision engine in Relite Smart Trader are shipping today on their own; Riddhi is the common foundation that carries such capabilities across the rest of the catalogue without each being rebuilt from scratch.
Which models does it actually use?
Open-weight models are the base — currently the Qwen family, with Gemma kept as a hedge so no single ecosystem is a dependency — served locally or on our own GPUs. External APIs, including frontier ones, are used for the small share of requests that genuinely need them. Which model handles a given request is the router's decision, not a setting you have to understand, and the whole point of the design is that the answer can change without anything else changing.
What happens when a model is unavailable or too expensive?
Every route has an ordered fallback chain, used when a backend is down, times out, overflows its context or would breach a budget. A chain is not allowed to silently jump more than one cost tier. A backend that keeps failing is taken out of rotation for a cool-down rather than retried into the ground, and the last resort for a conversation is a local model with an honest notice that capability is reduced. The platform degrades; it does not go blank.
How do you stop an agent being hijacked by a document it reads?
By never letting retrieved text act as instruction. Anything an agent reads or a tool returns is delimited and marked as untrusted data, and nothing found inside it can widen what the agent is allowed to do. Beyond that: tools are default-deny and allow-listed per agent, code runs sandboxed with no outbound network, side-effectful actions need human approval, and every tool call is validated against its schema before it is dispatched. Any one of these can be defeated in isolation, which is why there are several.
How do you know a change made it better rather than just newer?
An evaluation harness, treated as infrastructure rather than a later addition. Golden sets drawn from real customer questions, retrieval and agent scoring, and regression gates in continuous integration mean a model swap is judged on measurements against your work, not on a benchmark table. With a capable open model arriving every few weeks, being able to answer "is this better for our users?" within the hour is what makes riding that cadence possible at all.
Can it use our own fine-tuned model, or our existing AI provider?
Yes. Every component treats the model as a swappable commodity behind a common interface, and the API surface is the OpenAI-compatible one, so bringing your own weights or your own provider key is a routing-policy change rather than an engineering project.
When will it be available?
The plan runs in phases with explicit exit criteria rather than fixed dates, because a date met by shipping something that does not work helps nobody. The roadmap section above lists the phases in order. Ask us where it has actually reached — we will tell you plainly.

Talk to us about Riddhi AI

Tell us what you need and we'll come back to you — usually the same working day.

Other Products

Business Management System

SME back-office covering leads and clients, sales, income and expenses, double-entry accounting, and a full HR module with biometric attendance and payroll.

Explore
Enterprise Business Management Suite

Clean-architecture ERP spanning CRM, distribution, accounting, inventory, HR, engagement and reporting, with order-to-cash implemented end to end.

Explore
Relite Blood Bank

Blood bank and transfusion service management covering donor registration, collection, testing, component preparation, cross-matching, issue and statutory registers.

Explore
#
#
#
#
#
#
#
#
#