LedgerMind AI
A business recordkeeping and intelligence platform for informal bookkeepers such as small shop owners, contractors and farmers. It uses a template-driven recording engine in which AI only drafts records for user confirmation, numbers are computed by code from saved records, and the model is responsible only for generating narrative analysis.
Timeline (last 90 days)
LedgerMind AI positions AI output as draft records, which are saved only after user confirmation; the AI does not write directly to the database or confirm payments.
LedgerMind AI uses a template-driven record engine, where different business types share the same backend but each has its own fields, record types, and reports.
LedgerMind AI stores passwords hashed with PBKDF2-HMAC-SHA256, and session tokens store only hashed values.
LedgerMind AI has end-to-end functionality working for the restaurant and supermarket scenarios, including record entry, reports, the AI note parser, and the AI business analyst.
In business types such as contractors, pharmacies, farms, and logistics, LedgerMind AI is only partially usable, with issues such as some records not being counted and totals displaying as ₦0, and the overview page and report page potentially contradicting each other.
Outside the shop scenario, LedgerMind AI's AI extraction may misjudge the direction of funds (receipt or payment), store placeholder words as customer names, and not ask follow-up questions when the amount is unclear.
On the `/api/v1/ai/interpret-record` endpoint, Sentry recorded 7 HTTP 503s, 3 HTTP 429s, and 1 inference timeout, and the original retry logic waited only 1 second and 2 seconds, with all three attempts failing within 3 seconds.
A single business analyst request takes about 45 seconds, and the output is truncated at 2,500 tokens; the author plans to raise the output limit to 5,000 tokens.
LedgerMind AI's frontend uses React + TypeScript + Vite and is deployed on Render, its backend uses FastAPI + SQLAlchemy + Alembic + PostgreSQL and is also deployed on Render, its AI inference uses Google-hosted Gemma, and its monitoring uses Sentry.
LedgerMind AI's "profit" calculation is not yet implemented, and the author plans to answer profit questions with net cash flow (inflows minus outflows) and clearly label that this is not audited profit.
4 more timeline events
All Facts (14)
LedgerMind AI positions AI output as draft records, which are saved only after user confirmation; the AI does not write directly to the database or confirm payments.
50%UnverifiedLedgerMind AI uses a template-driven record engine, where different business types share the same backend but each has its own fields, record types, and reports.
50%UnverifiedLedgerMind AI stores passwords hashed with PBKDF2-HMAC-SHA256, and session tokens store only hashed values.
50%UnverifiedLedgerMind AI has end-to-end functionality working for the restaurant and supermarket scenarios, including record entry, reports, the AI note parser, and the AI business analyst.
50%UnverifiedIn business types such as contractors, pharmacies, farms, and logistics, LedgerMind AI is only partially usable, with issues such as some records not being counted and totals displaying as ₦0, and the overview page and report page potentially contradicting each other.
50%UnverifiedOutside the shop scenario, LedgerMind AI's AI extraction may misjudge the direction of funds (receipt or payment), store placeholder words as customer names, and not ask follow-up questions when the amount is unclear.
50%UnverifiedOn the `/api/v1/ai/interpret-record` endpoint, Sentry recorded 7 HTTP 503s, 3 HTTP 429s, and 1 inference timeout, and the original retry logic waited only 1 second and 2 seconds, with all three attempts failing within 3 seconds.
50%UnverifiedA single business analyst request takes about 45 seconds, and the output is truncated at 2,500 tokens; the author plans to raise the output limit to 5,000 tokens.
50%UnverifiedLedgerMind AI's frontend uses React + TypeScript + Vite and is deployed on Render, its backend uses FastAPI + SQLAlchemy + Alembic + PostgreSQL and is also deployed on Render, its AI inference uses Google-hosted Gemma, and its monitoring uses Sentry.
50%UnverifiedLedgerMind AI's "profit" calculation is not yet implemented, and the author plans to answer profit questions with net cash flow (inflows minus outflows) and clearly label that this is not audited profit.
50%UnverifiedLedgerMind AI's strategy of using an open-weight model (Gemma) provides a fallback: when the hosted endpoint returns unrecoverable errors, inference providers can be switched or self-hosting used, without waiting for a single vendor to recover.
50%UnverifiedThe author believes the next key step for open-weight models is fine-tuning to adapt to users' real ways of expression such as informal notes, monetary amounts, and fund-direction phrasing, where general-purpose APIs perform very poorly.
50%UnverifiedLedgerMind AI's browser voice input is based on the Web Speech API, with text input retained as a fallback.
50%UnverifiedLedgerMind AI explicitly states that its reports are not audit-grade accounting, and that reconciliation discrepancies are only review signals, not evidence of missing funds or misconduct.
50%