LeonardoSouza

Why Groq and not OpenAI directly?

The SDK is the same; only the base_url changes. Llama 3.3 70B on Groq delivers far lower latency at a cost a personal project can carry, and switching providers is one line of configuration if the trade-off changes.

Why truncate history at 10 messages?

The month's financial context already fills the system prompt. With no cap, a long conversation drives cost per request up and answer quality down — 10 messages covers the continuity of a finance chat.

Why a chronological split and not a random one?

Home prices move over time. A random split puts a 2009 sale in training and a 2007 sale in test, and the model ends up evaluated while knowing the future. The cut is by date, validation was the year 2009, and the final test ran against the 175 sales from 2010 — a year training never saw, with disjoint ids.

Why log the requests that fail with 422 as well?

Logging only successful predictions measures the model, not the service. Recording the 422s alongside them is what shows which contract the client is getting wrong and how often — that is where the 94.3% success rate over the report's 300 calls comes from, instead of an impression that everything was fine.

Why not retrain after the 2010 validation?

The model trained through 2009 came in at 9.32% MAPE on a year it had never seen, against 9.83% on the 2009 holdout — so there was no degradation to fix. Retraining with no sign of degradation swaps a measured model for a new, unmeasured one; the decision to leave it alone is written down in the validation report.

Why an outbox instead of calling the email provider inside the transaction?

An HTTP call inside a transaction holds a pool connection while it waits on the network. The outbox breaks that into three steps: a short transaction that claims the row, a send with no lock at all, and a short transaction that records the result. The Idempotency-Key is the row id, so a resend after a crash is a no-op.

Why does the rate limiter fail OPEN?

If Redis goes down, blocking all traffic takes checkout down with it — letting traffic through for a few minutes costs less than not selling. It logs ERROR so the failure raises an alert, and the circuit breaker keeps every request from paying the connection timeout while Redis is out.

Why a modular monolith and not microservices?

Solo project, ~100 orders/month. Microservices here would buy network latency and deployment complexity without solving a single problem I actually have. The modules have clear boundaries for the day splitting them pays off.

How do you import warranty data from a PDF without risking a bad extraction landing in a contractual document?

The import script never writes to the database. It emits a .sql you review as a diff, and it only becomes a migration after someone reads it; the check is structural (how many records, how many warranties, how many preventive items), because eyeballing hundreds of lines of text is not verification, it is hope.

Why does this go in the portfolio as a module and not as a whole product?

It is a team project: most of the repository's commits belong to another developer. What I claim is what the history confirms as mine — the homeowner's manual module and its optimizations. The rest of the system is here as context, not as authorship.

Why RAG and not fine-tuning?

The content changes with every game patch. Retraining a model on every change is expensive and goes stale fast; a curated corpus that gains new documents every day is current by construction, and the LLM only comes in to turn text into JSON and to answer grounded in what was retrieved.

Why does ingestion run on GitHub Actions instead of inside the API?

A serverless function has an execution time ceiling, and the daily ingestion (scraping, embeddings and LLM curation) does not fit under it. The cron job runs on Actions and only writes to the database; the API keeps reads and chat, which are fast enough.

Why truncate the embedding to 1024 dimensions?

The provider returns 3072 dimensions by default and pgvector's HNSW index has a practical limit below that. Matryoshka truncation (the dimensions parameter on the call itself) makes the vector fit the index without switching providers or giving up similarity search.