gharasathi: A Home Local AI Built on a Single Mac Studio
gharasathi runs AI locally on a Mac Studio, keeping family finance and photo data at home.
gharasathi is an AI system that runs entirely inside the home, named from the Marathi for "for the home." It integrates money, photos, and documents, with inference performed on a local Mac Studio, eliminating data-leakage risk architecturally. The system follows three principles: data stays home, code does the math while the model explains, and every finding shows evidence. The project began on a $200 mini PC centered on a Neo4j graph database with a 3B model, then upgraded to Qwen3.5-122B on a Mac Studio. The Hishob finance module reads bank data and bills to help families track where money goes, emphasizing the separation of deterministic parsing and model explanation. The roadmap also includes hosting the photo library locally with Immich to replace iCloud and Google Photos.
What Is gharasathi?
gharasathi (Marathi for "for the home") is an AI system that runs entirely inside the home. It brings together money, photos, and documents to help families make better decisions—without sending data to anyone's cloud.
The project is built on a core belief: home data should stay at home. Inference runs on a Mac Studio inside the house, eliminating the risk of data leakage at the architectural level.
Three Principles That Run Throughout
gharasathi follows the same rules across all modules, and these three principles form the design philosophy of the entire system:
- Data stays at home. Inference runs on a Mac Studio inside the home, with no reliance on external cloud services.
- Code does the math, the model does the explaining. No number comes from the model's "imagination"—the computation logic is handled by deterministic code.
- Every finding shows its evidence. That means a transaction record, or a quoted line from a document.
The third principle is especially critical. In a home finance context, any conclusion an AI produces must be traceable and verifiable; otherwise users can't trust it. gharasathi builds that trust by forcing the display of raw evidence.
Origins: Privacy-First AI on a $200 Mini PC
The project started on a $200 mini PC, centered on a Neo4j graph database paired with a small 3B-parameter model. That starting point makes one thing clear: home-grade AI doesn't require expensive hardware.
The article series documents several key technical decisions:
Running an LLM on 6GB of RAM
Model selection for edge AI is a trade-off problem under real-world constraints. The series' "Running an LLM on 6GB RAM: Model Selection for Edge AI" discusses in detail how to pick the right model within limited memory.
Personal AI Security Lessons from OpenClaw
Another piece, "What OpenClawAd Teaches Us About Personal AI Security," explores the security boundaries of personal AI assistants. Home AI handles the most sensitive data—bills, photos, family finances—so security design can't be an afterthought.
Neo4j is a native graph database that stores entities and their relationships as nodes and edges, rather than as tables and rows like a relational database. In gharasathi's case, home finance data is naturally graph-structured: a transaction links to a merchant, an account, and a bill; the bill in turn links to a family member or a spending category. Storing these relationships in a graph database makes multi-hop queries like "where did this money ultimately go?" direct and efficient, and it also lets the model trace a chain of evidence along the graph structure when explaining. For a home AI system that must show evidence, another advantage of a graph database is that every edge can carry source metadata, supporting verifiable reasoning paths.
The Finance Module Hishob: Local AI Helps You Find Where the Money Went
Hishob (हिशोब, meaning "accounts") is gharasathi's finance module, a private money dashboard. It reads bank data, bills, and statements, figures out where the money is going, and explains what to do about it.
The module's technical evolution is interesting:
From 3B to Qwen3.5-122B
The project initially ran a 3B model on a mini PC, then later ran Qwen3.5-122B on a Mac Studio as a home financial analyst. The jump in model capability brought by the hardware upgrade produced a qualitative change in analysis quality. But the project also emphasizes "keeping it honest"—keeping the model honest, meaning it must not generate numbers out of thin air.
The gap between 3B and 122B is not just a forty-fold difference in parameter count; it shows up in reasoning ability, instruction following, and long-context understanding. A small 3B-class model can run on edge devices with 6GB of RAM and is suited to classification, extraction, and simple Q&A, but it struggles with multi-step financial analysis, cross-bill correlation, and natural-language explanation—and it hallucinates more readily. A 122B-class model (like Qwen3.5-122B) needs far more VRAM and compute, and typically must be quantized to run on a Mac Studio's unified memory, but it performs markedly better at complex reasoning and at obeying constraints like "explain only, don't invent numbers." This leap in capability from the hardware upgrade is precisely the core motivation for migrating the project from a mini PC to a Mac Studio.
Reading Bills Without AI
A counterintuitive detail: one piece in the series, "Reading Bills and Statements Without AI, and the Money They Found," is about reading bills and statements without using AI, and the money problems discovered as a result. This reflects the project's pragmatic attitude—AI is a tool, not an end, and deterministic parsing is more reliable than a model in many scenarios.
Deterministic parsing means extracting fields from bills using rules, regular expressions, template matching, or structured-format parsers (such as CSV/PDF table parsers), rather than letting a language model "read" and generate the results. Its advantages are reproducibility, auditability, and zero hallucination: the same input always yields the same output, and when something goes wrong you can pinpoint exactly which rule failed. Its drawback is fragility in the face of documents with varying formats and poor scan quality, requiring maintained parsing rules for each bank or merchant. gharasathi's approach puts deterministic parsing first, with the model only responsible for explaining the structured results—preserving the flexibility of natural-language explanation while ensuring numbers are accurate.
What's Next: Taking the Photo Library Off the Cloud
The project roadmap also includes a photo module: deploying Immich on the Mac Studio to move the family photo library off iCloud and Google Photos. This is in keeping with the project's privacy-first philosophy throughout—photos are among the most private family data and likewise shouldn't be hosted on a third-party cloud.
Immich is an open-source, self-hostable photo and video management platform, often seen as a self-built alternative to Google Photos. It offers face recognition, object detection, smart search by location and time, album sharing, and more, with these capabilities performed by locally running machine learning models and the photos themselves stored on the user's own server or NAS. In gharasathi's architecture, deploying Immich on the Mac Studio means the photo library's storage, indexing, and smart retrieval all happen inside the home network, sharing the same "data doesn't leave the house" infrastructure as the finance module and reusing the existing local inference compute.
Implications for Home AI
The value of gharasathi isn't just that it's an open-source project—it demonstrates a viable architectural paradigm for home AI:
Local inference is a hard guarantee of privacy. Data not leaving the house isn't a promise made by a privacy policy; it's enforced by physical isolation.
Verifiability is the foundation of trust. The three rules—"code computes, model explains, evidence is shown"—solve the biggest trust obstacle for LLMs in finance: hallucinated numbers.
The hardware barrier is disappearing. From a $200 mini PC to a Mac Studio, the hardware range for home AI is getting wider, and ordinary families are fully capable of hosting local inference.
For developers who care about data sovereignty and want to run AI in their own homes, this series offers a complete reference path from hardware selection and model choice to concrete applications (finance, photos).
Related articles
Bilibili's Hit Dify Tutorial: A Practical Guide to Building 20+ AI Apps from Scratch
A distilled guide to Bilibili's top-rated Dify tutorial: Windows + Docker deployment, MySQL connection, five app types, model integration strategy, and publishing.
Agent Skills Beginner's Guide: From Concepts to Multi-Agent Collaboration in Practice
A comprehensive guide to Agent Skills: core concepts, components, the importance of Description, standardization, beginner learning paths, and multi-agent collaboration.
One Neuron Detects Both Cats and Cars? We Reproduced It and Found a New Explanation
The cat-car neuron reinterpreted: researcher Cammarata argues it distinguishes pointy from round, not cats from cars. We reproduced the experiment and found the new explanation holds on most inputs—but cars are the exception. The interpretability debate continues on language models.