ROMAN TEN · APPLIED AI / FULL-STACK

Answers
with sources.

I work across data ingestion, retrieval, backend APIs, and user-facing interfaces. WRS brings those pieces together in an assistant that answers from indexed website content.

Based in Suwon, South Korea

Explore WRS Email me Resume PDF

01 / 02Selected work

SELECTED WORK

From retrieval
to interface.

WRS is the main case study: the implementation is public, and you can try the website assistant above. The model training project below shows a separate evaluation workflow.

01 / FEATURED CASE STUDY

RAGFASTAPIPGVECTORJAVASCRIPT

WRS — a source-grounded assistant for websites

WRS turns website pages and uploaded content into a searchable knowledge base. It combines semantic and keyword retrieval, then returns answers with source links through an embeddable chat widget. The same service also supports private assistants and messaging integrations.

Try WRS Inspect the code

SYSTEM FLOW

CRAWLINDEXRETRIEVEANSWER

VECTOR WEIGHT70%

KEYWORD WEIGHT30%

Configured retrieval weights, not accuracy results

01 · CONTENT

Crawls static and JavaScript-rendered pages, chunks text, and indexes multilingual embeddings.

02 · RETRIEVAL

Combines pgvector similarity with PostgreSQL full-text search before composing an answer with source links.

03 · DELIVERY

FastAPI serves an embeddable website widget, private assistant views, and messaging integrations.

04 · FEEDBACK

A pilot analytics flow records low-confidence and unanswered questions, feedback, and response times for review.

The public repository documents the implementation. A public benchmark for answer quality or customer impact has not been published.

02 / MODEL ADAPTATION

QLORAMLX-LMEVALUATION

Qwen3 0.6B flashcard fine-tuning

Fine-tuned a 4-bit Qwen3 0.6B model with MLX QLoRA on 621 synthetic examples to generate JSON flashcards from a supplied passage. Published the adapter, training code, data, and evaluation workflow.

On 20 frozen passages, the adapter passed schema and exact-source checks in 20/20 cases, versus 14/20 for the prompted base model after removing its Markdown fences. Strict semantic review passed 10/35 separate family-held-out cases; all training and evaluation data are synthetic.

View model adapter Code, data & evaluation

Answerswith sources.

From retrievalto interface.

WRS — a source-grounded assistant for websites

Qwen3 0.6B flashcard fine-tuning

Answers
with sources.

From retrieval
to interface.