Back to projects

Research Engineering

HITL Multi-Evidence ESL Essay Assessment System

A local desktop research system for ESL essay assessment that combines linguistic feature extraction, local LLM rubric evaluation, and human review.

Python · Electron · Ollama · spaCy · NLTK · LanguageTool · Local LLM Inference

Status

Implemented and ready for experimental use. The associated research, "Beyond Prompt Engineering: A Human-in-the-Loop (HITL) Multi-Evidence Framework for ESL Essay Assessment," has been accepted for presentation at the 23rd AsiaCALL International Conference in Hanoi, Vietnam, November 14-15, 2026.

Overview

I am primarily responsible for building the technical system that operationalizes the research framework. The application combines essay ingestion, linguistic feature extraction, local LLM rubric evaluation, progress tracking, human review, final score entry, decision notes, and workbook export.

Problem

The research needed more than prompt experiments: it needed a working system that could process essays, preserve evidence, support human decision notes, and export completed workbook results.

Requirements

  • Run locally/offline instead of relying on cloud model inference
  • Process essay workbooks and preserve scoring progress
  • Extract linguistic evidence using NLP tools
  • Run local LLM rubric evaluation through Ollama
  • Let human reviewers compare evidence, assign final scores, and record decision notes
  • Export completed workbook results for experimental use

Architecture

01
Electron desktop frontend
02
Local HTTP API
03
Python backend
04
Linguistic analysis + local LLM inference
05
Human reviewer
06
Final score / notes / workbook export

Technical Decisions

Local-first architecture

The system uses a desktop frontend and local backend services so evaluation can happen without sending essays to cloud inference APIs.

Multi-evidence scoring

The workflow keeps linguistic features, model judgments, and human review visible as separate evidence sources instead of collapsing everything into one opaque score.

Human final decision

The system supports local LLM evaluation, but final scoring and decision notes stay in the human reviewer workflow.

Implementation

The system coordinates an Electron desktop interface with a local HTTP API and Python backend. The backend integrates spaCy, NLTK, LanguageTool, Ollama, and local Llama inference to produce evidence for human review. Reviewers can inspect essays, compare linguistic evidence with model judgments, enter final scores, document decision notes, and export a completed workbook.

Challenges, Tradeoffs, and Results

The main tradeoff is system complexity: local inference and multi-tool NLP integration require more orchestration than a cloud-only prompt prototype. The benefit is a research workflow that preserves evidence, supports human judgment, and can operate locally/offline. The human-vs-HITL experiment has not yet been completed, so the site does not claim accuracy, consistency, or efficiency improvements.