The science behind your desktop AI co-scientist
Nadhi pairs a frontier reasoning model with a desktop agent that has system-wide file access. It runs a five-persona multi-agent loop (Planner, Critic, Director, Synthesizer, and sub-agents that brainstorm to a final verdict across the RL, maths, or code approach you select), bulk-downloads 100+ open-access papers per Long-Horizon run from arXiv, PMC, bioRxiv and OpenAlex, indexes them for RAG, cross-references files across folders, and writes results back as format-preserved edits to .docx, .xlsx, .pdf and .pptx with a .bak undo. Below is the architecture that makes it reliable, and, honestly, what is shipping versus what is still in development.
Local Multimodal Agentic LLM: On-Device Intelligence
State-of-the-art open-weight model
Nadhi ships with a local multimodal agentic LLM, a state-of-the-art open-weight model optimized for on-device deployment. It achieves near-GPT-4 quality on reasoning, summarization, and instruction-following benchmarks while running entirely locally.
We use LiteRT (formerly TensorFlow Lite Runtime) for inference, enabling GPU-accelerated execution via Vulkan on NVIDIA and AMD hardware. The model runs in 4-bit GGUF quantization, requiring only 4-6GB VRAM for full-speed inference.
Bring Your Own Model (BYOM)
Specialized models for specialized needs
While the default local multimodal agentic LLM is exceptional for general-purpose research tasks, some specialized workflows benefit from domain-specialized models. Nadhi supports hot-swapping to any GGUF-compatible open-source model directly from the desktop UI.
| Model | Parameters | Specialty | Source |
|---|---|---|---|
| Local Multimodal Agentic LLM | - | General-purpose (default) | Bundled with Nadhi |
| Llama-3 | 70B | Advanced reasoning & logic | Meta |
| Mixtral | 8x7B | Multi-document synthesis | Mistral AI |
| Command-R | 35B | RAG & Web research | Cohere |
| Qwen-2.5-Coder | 32B | Data analysis & coding | Alibaba Cloud |
Hybrid AI Architecture
Combining Local Privacy with Cloud Scalability
Nadhi utilizes a Hybrid Architecture to deliver the best of both worlds. Sensitive clinical data and proprietary research are processed 100% locally via on-device models to ensure maximum privacy and regulatory compliance. Concurrently, public web research, literature queries, and intensive data aggregation can be seamlessly routed through our secure cloud infrastructure to leverage immense scalable computing power.
This includes deep support for complex multimodal diagnostics, such as ECG interpretation combined with blood report analysis. The local client orchestrates these tools seamlessly, ensuring sensitive patient records never leave the device while utilizing the power of a hybrid agent framework.
Agentic AI Architecture
Autonomous multi-tool task execution
Nadhi is not a chatbot. It's an autonomous agent. The LLM has access to a suite of tools and can chain them together to complete complex, multi-step scientific and clinical research workflows without human intervention.
Document Generator
Creates PDF summaries, DOCX manuscripts, XLSX tracking sheets from natural language instructions
Project Memory (RAG)
Semantic vector search across all uploaded project records. Ask questions about any project's history.
Web Research
Multi-round deep web search for market trends, competitor analysis, and latest scientific literature
Communication Gateway
Sends messages, documents, and reports via Telegram, WhatsApp, or Email automatically
File Management
Creates, organizes, and searches project folders. Converts handwritten notes to structured PDFs.
Scheduling Engine
Cron-based reminders and follow-up notifications via messaging channels
AI Co-Scientist Multi-Agent Architecture
Empirical discovery loops & long-horizon runs
Nadhi implements a complete AI Co-Scientist framework designed for empirical verification, paper writing, and data analysis. The core reasoning loops (/experiment and /longrun) behave as autonomous multi-agent pipelines managed by a robust safety-gate model.
1. Single Discovery Loop (/experiment)
Triggers an autonomous discovery turn: spawns specialized sub-agents that brainstorm to a final verdict on the hypothesis (across the RL, maths, or code approach you select), acts as a Planner to write code and pip-install dependencies, runs the script under a single secure user-approval click, Critiques outcomes, and Synthesizes a markdown research report on disk.
2. Long-Horizon Director Loop (/longrun)
Budget-bounded iterative runs: Prompts a Director agent to guide the run over multiple loops, deciding whether to refine previous results, branch to related hypotheses, or stop. Automatically compiles a detailed journal.md and lists novel observations as they accumulate over hours.
Core Co-Scientist Pipeline Primitives
✓ Dynamic Todos
Generates topic-specific, dynamic progress checklists at startup and ticks them off in real-time as tasks complete.
✓ Absolute Stop / Abort
Teardown signal plumbs through deep research streams, concurrent downloads, and host script executions instantly.
✓ Checkpoint Resume
Smart checkpointing on disk skips cached download steps, allowing interrupted long-horizon runs to resume seamlessly.
Native TUI Environment
Native Process Execution
Unlike generic AI tools that execute directly on your operating system with full unbounded access to your files, Nadhi runs its entire AI stack via a native environment on Windows utilizing strict agentic tool boundaries. The AI securely interacts with the system strictly through predefined tools, mitigating prompt injection risks.
Why this matters vs. OpenClaw & other agents
❌ Generic agents (OpenClaw, etc.)
- • Execute code directly on host OS
- • Full filesystem and network access
- • Can install packages, modify system files
- • No enterprise-specific controls
✅ Nadhi
- • Runs via robust agentic tool calls
- • Controlled access to host filesystem
- • Strict port-forwarded networking
- • Audit logging built-in
Roadmap: what ships today vs. what's next
We label the future as the future. No vaporware in the present tense.
Nadhi today is a single grounded research agent whose sub-agents brainstorm to a final verdict across the approach you select. The fully autonomous, evolutionaryhypothesis-ranking direction (à la DeepMind's AI Co-Scientist) is something we're actively building toward, but we will not claim it until it ships. Here is the honest split.
✅ Shipping today
- • Deep research that downloads hundreds of full papers + citations into your project
- • Retrieval-augmented synthesis across your whole corpus and adjacent fields
- • Source-traceable hypothesis, review & report drafting (you decide)
- • Multi-document & vision analysis (PDF, DOCX, XLSX, images, ECG/scans)
- • Quantitative simulation, Monte Carlo & on-device SAC forecasting
🛠️ In development, not yet available
- • Evolutionary hypothesis tournaments, specialised agents that generate, brainstorm, rank & evolve hypotheses over many generations
- • Autonomous computational-experiment loop, read a protocol, write the code, run it sandboxed, analyse, iterate (in-silico / dry-lab only)
- • Fully air-gapped on-device inference for institutions
These are research goals, stated as goals, not current capabilities.