Rahul Roy

Synchronous web scraping and large PDF extraction block API event loops, suffer from memory spikes under concurrent requests, and fail silently when upstream rate limits hit.

Decoupled ingestion into an asynchronous FastAPI gateway backed by Redis message brokering, Celery worker pools, Playwright headless browsers, and structured Gemini AI extraction—containerized with Docker and Kubernetes HPA manifests.

Sending raw CSV rows directly to an LLM causes arithmetic hallucinations on aggregations, blows past token context windows, and leaks sensitive tabular records.

Extracts only column headers and dtypes for the Groq LLaMA-3.3-70B prompt, synthesizes executable Python/Pandas code, validates the syntax tree via a custom AST Security Visitor, and executes deterministically in a restricted local namespace.

Passive re-reading of technical notes leads to rapid memory decay, while manual flashcard creation is tedious and standard LLM wrappers block the UI waiting for full JSON payloads.

Streams flashcards in sub-second NDJSON chunks from Groq LLaMA-3.3-70B and schedules long-term reviews using a deterministic SuperMemo SM-2 algorithm updating card easiness factor (EF >= 1.3), repetition count, and day intervals.

Symmetric GARCH models ignore the financial leverage effect where negative news shocks amplify volatility far more than positive shocks, and random K-fold splits leak future regimes into training.

Fits Nelson's asymmetric log-variance EGARCH(1,1) model via Maximum Likelihood alongside a tri-level sentiment cascade (GDELT v2, Google Trends PCA, VADER) and a chronological 80/20 Random Forest regime classifier across 17 years of Nifty 50, Sensex, and Bank Nifty data.