Daniel A. F

A four-agent pipeline that turns a research topic into a sourced, formatted document. Instrumented end to end: cost per document, per-agent attribution, and a deterministic eval suite that gates the output.

770,000 reviews embedded and indexed in nine days, 500,000 of them name-resolved by a backfill that verifies its own join before writing. Retrieval runs five stages: corpus grounding, LLM signal extraction, multi-vector pgvector search over HNSW, a quality gate that can refuse to answer and ask instead, and psychographic reranking. A separate four-agent loop simulates how a specific user would rate and review an item they have never seen, with a critic that rejects drafts and sends them back. Scored offline on stratified holdout sets with NDCG@10, Hit@10 and MRR across six ablation variants, one of which beat the full pipeline and is written up as such.

Configurable ETL pipeline automating data cleaning, feature normalisation, and class-imbalance handling (SMOTE/ADASYN). Dual-interface (CLI & REST API) exposes the full ML lifecycle — ingestion to model serialisation.

Flask wrapper that dynamically routes requests to arbitrary inference functions with schema-based input validation. HTML generation engine parses Python dicts to construct semantic forms — abstracting the frontend layer for ML engineers.

Low-latency Rust backend for a PvP auto-battler. Integrates Solana wallet auth and handles complex character state management optimised for concurrent game sessions.