UnoOne LEAF is InBharat's new direction for building a Local Evolving Agent Fabric (AGI) designed for India. It's a local-first agent OS that analyzes data, acts through tools, verifies results, remembers failures, and improves skills over time, addressing the unique challenges of India's diverse digital landscape.
A practical, jargon-free guide for Indian engineering teams and founders — part of the Learn AI with Reeturaj series on InBharat AI.
Imagine an AI that truly understands your world, not just a cloud server thousands of kilometres away. That's the vision driving UnoOne LEAF, our new direction for building a Local Evolving Agent Fabric right here at InBharat. We're not just chasing the global AGI dream; we're building it for India, starting with your pocket.
Most AI today lives in massive data centres, far removed from the daily realities of an Indian user. It struggles with regional languages, slow 4G networks, and the diverse data formats we encounter – from handwritten Sahayaak Seva forms in a remote clinic to a student's resume in a Tier-2 city. Cloud-dependent AI often means latency, high data costs, and a lack of true privacy. We need an AI that's always on, always learning, and always local.
My team in Chennai spent months seeing how existing models faltered with nuanced Indian contexts. A simple document extraction task, for instance, often needed extensive re-engineering for local dialects or specific government form layouts. This isn't just about translation; it's about cultural and contextual understanding.
UnoOne LEAF is designed as a local-first agent operating system. It's built to analyze any data on your device, act through tools, verify its own results, remember its failures, and continuously improve its skills over time. Think of it as a personal, evolving intelligence living directly on your phone or local device.
Here's how we're structuring it:
flowchart TD
A["UnoOne Android App"] --> B["Voice / Camera / Screen / Files"]
B --> C["Local LEAF Core"]
C --> D["Qwen Router"]
D --> E["Skill Engine"]
E --> F["Tool Layer"]
F --> G["Verifier Engine"]
G --> H["Memory Ledger"]
H --> I["Sleep Learning Loop"]
I --> E
At its heart, the Local LEAF Core orchestrates everything. The Qwen Router acts as the brain, directing tasks to the right Skill Engine. Our Tool Layer allows the agent to interact with the device and external services, much like a human uses different apps. Crucially, the Verifier Engine checks the agent's work, and the Memory Ledger records every success and failure. This feedback loop, powered by the Sleep Learning Loop, is how LEAF gets smarter every night, without needing constant cloud updates.
We're starting with a Qwen-first model plan, leveraging smaller, efficient models like qwen2.5:1.5b for tasks like intent routing and JSON planning, and qwen2.5-coder:1.5b-instruct for code generation and tool repair. This allows us to keep the intelligence local and nimble, even on mid-range Android devices common across India.
Building for India means designing for constraints that become strengths. Our focus on local models means less reliance on constant internet connectivity, crucial for users in areas with patchy 4G or limited data plans. This offline-first approach significantly cuts data costs (in ₹) and ensures privacy, as sensitive data never leaves the device.
The architecture is modular. The Android app, already familiar to UnoOne users, acts as the user-facing agent. It communicates with the leaf-core, which can run locally on the device or even a nearby low-power server. This flexibility is key for deploying across diverse environments, from a farmer's smartphone to a small business's local server.
Our initial feature, the Any Data Analyzer, is designed to handle the messy, unstructured data prevalent in India – PDFs, text, screenshots, CSVs, even app logs. It will extract key entities, identify missing information, and suggest actions, all locally. This is a practical step towards making AI useful for everyday tasks, like processing a Sahayaak Seva health form or analyzing a university profile for UniAssist.
UnoOne LEAF isn't just an upgrade; it's the foundational intelligence layer for multiple InBharat products:
This local AGI fabric will allow us to deploy intelligent features that are deeply integrated with the user's context, respecting their data and working seamlessly regardless of network conditions. It's about building practical AI that truly serves India's diverse needs.
Q1: How will UnoOne LEAF handle regional languages and dialects?
A1: Our local Qwen models are being fine-tuned with diverse Indian language datasets. The Skill Engine and Memory Ledger are designed to learn and adapt to specific linguistic nuances and cultural contexts over time, improving accuracy for regional languages without relying on a distant cloud service.
Q2: What kind of devices will UnoOne LEAF run on? A2: We are optimizing for mid-range Android smartphones, which are widely accessible in India. By using quantized Qwen models (0.5B / 1.5B) and a modular Python kernel, we aim for efficient performance on devices with 4GB RAM or more, ensuring broad accessibility.
Q3: How does the "Sleep Learning Loop" work without internet?
A3: The Sleep Learning Loop operates entirely on the local device. It reviews past task failures stored in the Memory Ledger, identifies patterns, and then intelligently mutates and tests variations of skill prompts. The best-performing skill variants are then promoted for future use, all without needing external data or cloud connectivity.
Building true AGI for India means starting local, embracing constraints, and focusing on practical, on-device intelligence. UnoOne LEAF is our commitment to this vision, creating an evolving, self-improving AI fabric that understands and serves the unique needs of Bharat. Explore more about our practical AI solutions at InBharat.ai or dive into how we're building AI for Indian Languages and optimizing AI for low-resource environments.
#InBharatAI #LocalAGI #MadeInIndia #AIForBharat