Kharcha · ghar ka hisaab, phone se bahar nahi jaata

A tiny, private expense ledger for my parent. Paste any Indian bank/UPI SMS, or say "aaj sabzi wale ko 80 diye", and a 4B open-weight model fine-tuned on Tinker turns it into a categorised ledger row. The model runs offline on a laptop with Ollama; bank SMS never leave the machine. Backboard remembers the corrections ("Sharma Kirana is groceries", "Rahul is my son") and answers Hinglish questions about the month. The demo is hosted on Render: https://kharcha-4aax.onrender.com (free tier, ~1 min cold start).

Built for the DEV Hacktoberfest Weekend Challenge "Build for a Friend" (Oct 2026).

SMS / voice note ─▶ Qwen3.5-4B + Kharcha LoRA ─▶ {"direction","amount","counterparty","channel","account_last4","date","category"}
                          ▲ hints                         │
                 Backboard memory  ◀── corrections ───────┘

• data/ – program-labelled dataset generator (real HDFC/SBI/ICICI/Axis/Kotak/PNB/PhonePe/GPay/Paytm formats + Hinglish voice notes). Labels are exact because the program that writes the SMS also writes the label.
• train/sft_tinker.py – LoRA SFT on Tinker (Qwen/Qwen3.5-4B, rank 16, ~100 steps, < $1).
• train/eval.py – 150 held-out messages vs zero-shot Qwen3.5-4B and gpt-oss-120b. Produces out/results.md.
• train/export_adapter.py – download LoRA → PEFT → merge → GGUF → ollama create kharcha.
• app/ – FastAPI + single-page UI. PARSER_BACKEND=ollama (private) or tinker (hosted demo).

brew install uv ollama
uv venv --python 3.12 && uv pip install -r requirements-train.txt
cp .env.example .env            # add TINKER_API_KEY, BACKBOARD_API_KEY

.venv/bin/python data/gen_dataset.py          # 1200/150/150
.venv/bin/python train/sft_tinker.py          # prints KHARCHA_SAMPLER_PATH -> put in .env
.venv/bin/python train/eval.py --systems base,big,lora
.venv/bin/python train/export_adapter.py      # -> ollama model "kharcha"
.venv/bin/python train/eval.py --systems ollama

ollama serve &
PARSER_BACKEND=ollama .venv/bin/uvicorn app.main:app --reload   # http://127.0.0.1:8000

See out/results.md after running the eval (table reproduced in the DEV post).

• The adapter is ~150 MB and ours. Swap the base model by changing one string.
• Fine-tune cost under a dollar; inference cost at home is zero.
• A parent's bank SMS are the most private text they have. They stay on the laptop.