Korzystając z Large Language Modele dla Context-Aware Expense Tagging
Single intensive sesja z follow-up Q&A within 14 days
"Finally a ustrukturyzowany approach że handles edge cases w expense classification - the AI logic covered categories I hadn't even considered."
"Worked poprzez the programme alongside my day job. The pacing was realistic i the mentor feedback was specific, nie generic."
"The categorisation modele my stworzony są now running w production. That's a result I didn't expect ten early."
Standard classifiers look at a transaction description i assign a category. Że works dla straightforward cases. But co o a płatność do a venue że could być a zespół lunch, a client event, lub a conference room hire? Bez context, dowolne model będzie guess. Z context, an LLM może reason poprzez it.
Ten sesja jest dla people kto already have a working categorization pipeline i want do handle the 15 do 20 percent transactions że keep landing w the wrong bucket. My look at jak do pass contextual signals do an LLM: the merchant category code, the amount, the day week, nearby transactions w the same batch, i dowolne notes the submitter added.
Prompt engineering dla finanse tasks
My go poprzez prompt design w detail. The goal jest a ustrukturyzowany output, nie a conversational response. Ty want the model do return a category, a confidence signal, i a brief reason. My cover function calling z the OpenAI API i narzędzie używać z Claude do get consistent JSON back rather than free text że needs parsing.
My also talk o koszt. Running every transaction poprzez a frontier model jest expensive. The practical approach jest a tiered system: a fast cheap classifier handles the obvious cases, i the LLM tylko sees the ambiguous ones. My build że routing logic together.
Prerequisites
Ty should have Python experience i some familiarity z API calls. Prior exposure do prompt engineering jest helpful but nie Wymagane.
Co the path looks like
- Gdzie simple classifiers fail i dlaczego context jest the missing piece
- Contextual signals worth passing do an LLM: MCC codes, amounts, timestamps, notes
- Prompt design dla ustrukturyzowany categorization output
- Function calling i narzędzie używać dla consistent JSON responses
- Building a tiered routing system: fast classifier plus LLM fallback
- Koszt estimation i optimization dla production używać
- Logging i evaluating LLM categorization decyzje
Practical application
Each stage involves hands-on work z real expense datasets. Categorisation modele są stworzony incrementally, so Ty understand the logic behind each decyzja rather than copying a finished result.
- Labelling strategies dla ambiguous transactions
- Handling multi-currency i cross-border entries
- Validation pipelines do catch misclassification early
- Iterative refinement oparty na model output przegląd
Mentorship rhythm
Sesje są ustrukturyzowany around Twoje actual work, nie a fixed curriculum. Bring a specific categorisation problem i leave z a working approach, nie just theory.
"The feedback loop here jest tight enough że mistakes get caught przed they compound." - Participant, cohort 7