Runs offline · no API key

Bank reconciliation that never leaves your laptop.

Match statements to ledgers, catch the anomalies a spreadsheet misses, and export an audit-ready report where every row explains itself. Deterministic first — a local model only picks up what the rules can't.

4
pipeline stages
165
tests, fully offline
3
input formats
0
bytes sent off-machine

The pipeline

Four stages. Click one.

Each stage hands only its leftovers to the next. That's the whole design: the cheap, explainable work happens first, and the model never touches a row the rules already resolved.

Ingest

Loading.

Why it's built this way

Deterministic beats model, every time.

Rules learned the hard way while building it — and each one shows up in the architecture, not just the README.

Client data never leaves the machine

Extraction, embeddings, and match confirmation all run against a local llama.cpp server. There is no default network call to any provider, and no API key to configure. Hosted backends exist behind a Protocol if you want them — you have to opt in.

Fails loudly

A wrong number in a financial report is worse than no report. Extraction refuses rather than publishing figures it can't reconcile against the running balance.

Every row is traceable

The Matched tab carries a corroboration column: shared reference, matching description, or amount-and-date-only — so a reviewer knows which pairs to double-check.

Degrades gracefully

Kill both model servers and it still reconciles. Stage 1's matches are never lost to a Stage 2 failure — the leftovers just land in the review queue.

One bank at a time

PDF templates are built and validated per bank, not as a generic parser. Each real-world layout quirk gets locked into a regression test.

Download

Unzip and run. That's it.

No Python, no terminal, no dependencies. Download the zip for your platform, unzip it, and double-click the app. The AI models are bundled — nothing downloads on first launch.

Unzip the archive and run ReconciliationEngine\ReconciliationEngine.exe. On first launch it loads the AI models onto the GPU — this takes around 15 seconds. Nothing leaves the machine.

Questions

The ones people actually ask.

Does it work on Windows?
Yes. The engine is pure Python — FastAPI, pandas, openpyxl, pypdf, httpx all ship Windows wheels, and llama.cpp has native Windows builds. Only the README's shell commands need translating: py -3 -m venv, .venv\Scripts\Activate.ps1, and two separate PowerShell windows instead of backgrounding with &.
Do I need a GPU?
No, but a 3B model on CPU is slow — the default request timeout is 300 seconds for exactly that reason. On Windows with an NVIDIA card, a CUDA-enabled llama.cpp build is a large speedup. You can also skip the model servers entirely and get deterministic reconciliation only.
Which banks are supported?
HDFC today. PDF templates are built one bank at a time and validated against a real statement each, rather than shipping a generic parser that quietly misreads columns. Adding one means implementing build_prompt() and parse_response() and registering it — it then appears in the UI dropdown automatically.
Scanned PDFs?
Not in v1. Extraction reads the PDF text layer, so a scanned page yields nothing and raises a clear error instead of silently returning empty results. Scanned support needs OCR or a hosted vision model — both are wired but off by default.
Is my data sent anywhere?
No. The default configuration has no API key and makes no outbound call — both model servers run on localhost. The test suite is offline too, exercised against an in-memory HTTP transport.