Hishob
The local finance module of gharasathi. It reads bank data, bills and statements, analyzes the flow of funds, and provides verifiable explanations.
Timeline (last 90 days)
Hishob's nightly pipeline is roughly: at 01:30 read and archive new bills; every hour run about twenty deterministic detectors when data changes; at 02:30 the AI analyst reads the code-built fact pack and writes findings, which then go through fact-checking and editing.
Hishob's core design principle is "code computes, model explains": all totals, averages and matching are computed by code, the model is only responsible for finding correlations, explaining causes and suggesting actions, and the model never does money addition.
Hishob includes a fact-checker that rejects any AI finding containing amounts or percentages that cannot be traced to a data source.
Developer Koustubh published an open-source household finance dashboard project called Hishob on Dev.to, which runs on the home LAN and uses a local AI model to automatically analyze spending and give suggestions.
Hishob uses a mixed Go, Python and React tech stack, is deployed on a k3d (a lightweight Kubernetes tool that runs k3s in Docker) cluster, and uses a Neo4j graph database and Postgres for core data storage.
Hishob contains five microservices named in Marathi: aapla-dhan (Go, bank CDR sync), aapla-patrakar (Go, reads bills and statements), aapla-olakh (Go, PDF parsing and arithmetic verification), aapla-hushar (Python/FastAPI/LangGraph, rule detection and AI analysis), aapla-hishob (React, frontend dashboard).
Only two things in Hishob need to go out to the network, and both are read-only: the bank CDR interface and Gmail IMAP; all other inference is done locally on the Mac Studio.
Hishob's spending calculation rules are: transfers between accounts and transfers into the offset account do not count as spending, mortgage principal counts as savings (increasing equity), interest counts as a real cost, and refunds are automatically matched to the original payment rather than counted as extra income.
Each bank data source in Hishob syncs independently, and the dashboard only treats data as complete up to the last transaction of the oldest account, avoiding lagging data sources making the month's spending look too low.
Hishob is not fully automatic: bank transactions are manually triggered to sync once a week, bills are forwarded to a dedicated mailbox and then tagged by type by Gmail filters, and the nightly job only reads these tags, with other files placed in a Nextcloud folder.
4 more timeline events
All Facts (14)
In the author's data, in the months with the highest offset account balances, daily interest fell by about one fifth.
60%UnverifiedDeveloper Koustubh published an open-source household finance dashboard project called Hishob on Dev.to, which runs on the home LAN and uses a local AI model to automatically analyze spending and give suggestions.
50%UnverifiedHishob uses a mixed Go, Python and React tech stack, is deployed on a k3d (a lightweight Kubernetes tool that runs k3s in Docker) cluster, and uses a Neo4j graph database and Postgres for core data storage.
50%UnverifiedHishob contains five microservices named in Marathi: aapla-dhan (Go, bank CDR sync), aapla-patrakar (Go, reads bills and statements), aapla-olakh (Go, PDF parsing and arithmetic verification), aapla-hushar (Python/FastAPI/LangGraph, rule detection and AI analysis), aapla-hishob (React, frontend dashboard).
50%UnverifiedOnly two things in Hishob need to go out to the network, and both are read-only: the bank CDR interface and Gmail IMAP; all other inference is done locally on the Mac Studio.
50%UnverifiedHishob's spending calculation rules are: transfers between accounts and transfers into the offset account do not count as spending, mortgage principal counts as savings (increasing equity), interest counts as a real cost, and refunds are automatically matched to the original payment rather than counted as extra income.
50%UnverifiedEach bank data source in Hishob syncs independently, and the dashboard only treats data as complete up to the last transaction of the oldest account, avoiding lagging data sources making the month's spending look too low.
50%UnverifiedHishob is not fully automatic: bank transactions are manually triggered to sync once a week, bills are forwarded to a dedicated mailbox and then tagged by type by Gmail filters, and the nightly job only reads these tags, with other files placed in a Nextcloud folder.
50%UnverifiedHishob uses a fixed "small ontology" in the Neo4j knowledge graph, limiting node types and the relation vocabulary and not allowing generic LINKED_TO relations, which makes the local model's Cypher queries more reliable, and every Fact can be traced back to the source PDF.
50%UnverifiedThe Hishob dashboard generates a priority-sorted list of findings every morning, partly from deterministic rules (price increases, unpaid bills, subsidy gaps) and partly from the nightly AI analyst, with each finding carrying two action buttons, Explain and Resolve.
50%UnverifiedHishob's core design principle is "code computes, model explains": all totals, averages and matching are computed by code, the model is only responsible for finding correlations, explaining causes and suggesting actions, and the model never does money addition.
50%UnverifiedHishob includes a fact-checker that rejects any AI finding containing amounts or percentages that cannot be traced to a data source.
50%UnverifiedHishob's nightly pipeline is roughly: at 01:30 read and archive new bills; every hour run about twenty deterministic detectors when data changes; at 02:30 the AI analyst reads the code-built fact pack and writes findings, which then go through fact-checking and editing.
50%UnverifiedHishob is gharasathi's finance module, a private money dashboard that reads bank data, bills and statements, finds where the money goes and explains what to do about it.
50%