Skip to content
Mirun Kaushik
← All work

05Internship assignment2026

Transaction Pipeline

Cleans messy bank CSVs, flags anomalies, categorises with an LLM

My part
Solo
Status
Built
Stack
Python, FastAPI, Celery, Redis, PostgreSQL, Docker, OpenRouter
Code
GitHub ↗

Fig. — How data moves through it

  1. 01Dirty CSV
  2. 02Celery jobRedis broker
  3. 03Clean + dedupe
  4. 04Anomaly rules3× median · currency
  5. 05LLMuncategorised only(model call)
  6. 06Postgres + summary
Model call Rules, data, queues

01Problem

Bank CSV exports are messy: different column names, mixed date formats, currency symbols in amounts, duplicates and missing categories.

02Approach

Upload returns a job ID straight away and a Celery worker runs the pipeline in stages. Plain rules do the cleaning and the anomaly checks; the LLM is used only for transactions that are still uncategorised, and for the final summary.

03What I built

  1. 01Built it with AI coding tools; the design decisions were mine.
  2. 02A cleaning stage that maps column-name variants to one schema, normalises dates and amounts, fills defaults and removes duplicates.
  3. 03Two anomaly rules: amounts above 3× the account’s median, and USD charges at Indian-only merchants such as Swiggy, Ola and IRCTC.
  4. 04Batch LLM categorisation for uncategorised transactions, and a written summary of spending behaviour.
  5. 05Job tracking through pending, processing, completed and failed, with a results endpoint that never returns partial data.