Arure/Services/in production.

AI agents, in production.

Multi-agent systems, retrieval-grounded LLMs, computer vision, and document automation — shipped with eval harnesses on every agent action. Live in our own products (ThriveBase, TaxGPT) and in client deployments across the Gulf and Pakistan.

PatternsAgents · retrieval · vision · documents·StackAnthropic · OpenAI · Bedrock · SageMaker·DisciplineEvals · monitoring · cost-per-call

/01 — What we cover

/01

Multi-agent systems

Orchestrator agents dispatching specialised sub-agents — outreach drafting, next-best-action, deal diagnosis. Same pattern powering ThriveBase, available for your domain.

/02

Retrieval-grounded LLMs

RAG pipelines on your corpus — policy docs, regulations, product catalogues, ticket history. Citation-backed answers with the audit trail compliance asks for.

/03

Document automation

Invoices, contracts, claims, KYC packs — extraction, classification, routing, and approval workflows evaluated on your own corpus, not a vendor benchmark.

/04

Computer vision

Quality control on production lines, surveillance with anomaly detection, document OCR, and visual search — deployed on edge devices where latency or bandwidth demands it.

/05

Predictive modelling

Demand forecasting, churn scoring, fraud detection, lead scoring — XGBoost when it wins, deep learning when it earns its place. No model trained without a baseline.

/06

Eval & monitoring

Every agent action runs through an eval harness before shipping and a confidence/cost dashboard after. Hallucination tracking, drift alerts, cost-per-call by route. AI you can run with the same discipline as any other production system.

Anthropic

OpenAIAWS BedrockSageMakerAzure OpenAI

Hugging Face

PyTorchLangChainLlamaIndexPineconepgvector

/02 — What we ship

LiveTaxGPT in production · UAE tax assistant

Multi-agentThriveBase CRM with agents as the operating layer

MCP-nativeOdoo MCP server published on apps.odoo.com

Eval-firstEvery agent action measured before it ships

/03 — How it runs

/01

Use-case diagnosis

Two-week paid discovery. We map the workflow, score the data, and tell you whether AI is the right tool — or whether a deterministic rule would solve it cheaper. We say no when we should.

/02

Eval harness first

Before we write the agent, we write the test set. Golden examples, edge cases, and the metric we'll judge ourselves against. No agent ships without a baseline.

/03

Agentic build

Orchestrator + sub-agents, RAG, tool calls, guardrails. Built in your repo, your cloud, against your data — not on a vendor's hosted playground.

/04

Operate & improve

Production deployment with cost, latency, and hallucination dashboards. Weekly evaluation runs. The same engineering bench that built it stays on to tune it.

/04 — In production

← Scroll case studies →

Live · in-house venture · UAE

TaxGPT — retrieval-grounded UAE tax assistant

Arure Technologies

Challenge

UAE tax compliance moves quickly and authoritative guidance is scattered across FTA publications. Businesses needed conversational access to the rules without paying advisor fees for every question.

Solution

Retrieval-grounded LLM over a curated FTA corpus. Every answer carries citations back to the source paragraph. Confidence scoring routes uncertain queries to a fallback path instead of inventing. Hallucination tracking runs nightly against a held-out eval set.

Natural-language Q&A on UAE VAT and Corporate Tax
Citations on every answer — no answer without a source
Confidence-routed fallbacks instead of confident wrong answers
Live in production at /products/taxgpt

Live · in-house venture · multi-agent CRM

ThriveBase — agents as the operating layer

Arure Technologies · ThriveBase

Challenge

Every major CRM treats AI as an add-on layer for summarisation or autocomplete. The brief was to architect a CRM where agents are the operating layer — reading pipeline state, executing follow-ups, and proposing next-best-actions across deals, contacts, and accounts.

Solution

Multi-agent architecture on Anthropic + OpenAI. An orchestrator agent dispatches specialised sub-agents for outreach drafting, deal-state diagnosis, and meeting prep. Real-time pipeline data flows through retrieval, enrichment, and policy layers before reaching the user. Eval harness measures drafted-outreach acceptance per agent and tunes weekly.

Multi-agent architecture — not features bolted onto a legacy schema
Agents propose next-best-action on every open deal in real time
Acceptance-rate metrics per agent, tuned weekly against evals
Live read on thrivebase.app — built on the same bench as government work

Anthropic

OpenAI

Next.js

TypeScript

PostgreSQLpgvector

Live · marketplace · Odoo

AI Tool Calls via MCP — Odoo as an agent surface

Arure Technologies · Odoo marketplace

Challenge

ERP and AI live in separate worlds. Custom integrations break with every Odoo upgrade. We wanted a standard protocol surface so agents could work against Odoo the way they work against any other MCP server.

Solution

Published an MCP server for Odoo. Agents can read/write records, trigger workflows, and call out to Slack, Notion, GitHub, Drive, Postgres, and the wider MCP ecosystem — all governed by Odoo's existing permission model.

Live on apps.odoo.com — Odoo's official marketplace
Standard MCP protocol — works with Claude, ChatGPT, and any MCP client
Permission-model integration — no bypass of Odoo's access control
Foundation for agentic workflows on top of every Arure Odoo deployment

/ Other practices

ERP Implementation→Digital Transformation→Cloud & Data→

Start a project

Tell us what you’re building. We’ll tell you whether we’re the right team.

hello@arure.tech ↗

AI agents, in production.

PatternsAgents · retrieval · vision · documents·StackAnthropic · OpenAI · Bedrock · SageMaker·DisciplineEvals · monitoring · cost-per-call