23 Posts getaggt mit "devlog"

June 18, 2026, marks the triumphant completion of the 1st semester of our university capstone engineering project. From our humble beginnings as a simple Android convention scraper in March, Maeumjigi has evolved into a full-scale, clinical-grade AI mental health counseling ecosystem.

Our final sprint focused on two mission-critical objectives:

1. The Dual-Layer Safety Wrapper: Engineering a bulletproof guardrail architecture to guarantee zero-hallucination crisis intervention.
2. Production Cloud Serving & Final Defense: Sizing AWS EC2 GPU infrastructure budgets and delivering our final faculty presentation.

The final week of May 2026 (May 27–30) proved to be the grand culmination of our experimental modeling phase. After weeks of hyperparameter tuning, GPU out-of-memory crashes, and dataset revisions, our 4th Fine-Tuning cycle (fine_tuning_4) crossed the finish line.

This post reveals the technical breakthrough that resolved the chronic "over-counseling" trap, quantitative validation of our golden weight Checkpoint-4400, and the end-to-end system architecture designed for real-world deployment.

On May 21st and 22nd, the Maeumjigi Capstone team embarked on massive data generation and preparation for the 4th Fine-Tuning (fine_tuning_4).

Over the two days of May 19th and 20th, the automation of evaluation for the 3rd fine-tuning (fine_tuning_3) and the enhancement of backend/frontend features for building the actual app service proceeded in parallel.

Mid-May 2026 was supposed to be a calm interval after wrapping up our 2nd fine-tuning (fine_tuning_2). Instead, May 17th hit our team with a relentless double challenge: condensing hundreds of raw inference metrics into an automated technical report, while simultaneously battling a cryptic HTTP 401 Unauthorized cascading error across our newly integrated FastAPI and PostgreSQL backend.

Here is the complete post-mortem of how we automated our reporting and systematically neutralized the backend authentication breakdown.

On May 13th, after fixing the hardcoding bugs discovered yesterday, we finally resumed proper inference testing for the 2nd round fine-tuned models.

The 2nd round of fine-tuning (fine_tuning_2) parallel training, which was ambitiously started to correct the LoRA hyperparameter mistake (alpha/rank setup error), finally ended on May 12th!

May 8th was the day we were building our capstone design mid-term presentation (PPT) based on the fine-tuning results. However, while happily filling in the slides, I discovered a massive, chilling mistake.

May 7th was the day we synthesized the fine-tuning results of our 5 models (Llama3.3-70B, DSR1-70B, HCX-Omni-8B, Qwen35-9B, HCX-Text-1.5B) and proceeded with full-scale visualization and local environment setup.

On May 6th, we finally escaped the long swamp of fine-tuning! After several nights of work and hyperparameter tuning, the training for mid-size models like HCX-Omni-8B and Qwen35-9B was successfully completed.

After experiencing the OOM (Out of Memory) hell during the fine-tuning of massive 70B models, on May 4th, we attempted a bold change in strategy. By drastically reducing the batch size, we barely managed to push the 70B models (Llama3.3, DeepSeek-R1) to checkpoint-1200 and 1300, but the training time was absurdly long, resulting in terrible efficiency.

On May 3, having gained enough experience with the 8B models, our team finally challenged the ultimate kings of performance: fine-tuning massive 70B (70 billion parameters) models. We attempted to train both Llama 3.3 70B and DeepSeek-R1 70B, but as expected, we hit a massive hardware resource wall.

April 19 is the day we evaluated the performance of the model (M6) trained on the virtual synthetic dataset (v5) we painstakingly created, and wrote the 7th-week weekly report based on it. Was the data we generated while enduring a $10 bill bomb worth it?

April 14, 2026, was a masterclass in distributed pipeline fault tolerance. Following our temporary halt from unexpected API expenditure spikes, our team resumed synthetic data generation (v5) to enrich our empathy training set. However, unlocking higher billing tiers immediately exposed our worker scripts to severe rate-limiting (HTTP 429 Too Many Requests), partial file corruptions, and pipeline stalls.

Here is how we re-engineered the synthesis architecture into a self-healing, unmanned autonomous pipeline.

On April 9, using the Colab notebook we ambitiously completed yesterday, we finally attempted the first training (fine-tuning) of our artificial intelligence model (Llama 3.1). However, reality was not that easy, and we had to face numerous errors. It was also a tough day as we had to simultaneously create the presentation materials for the 6th week of the Capstone Design project.

April 8th was the day we experienced the biggest crisis and a 'laugh-and-cry' happening while working on the Maeumjigi project. It was the Claude API monthly limit exceeded & bill bomb incident.

April 7 was the day we attempted integration by running the frontend code built by a teammate in a local environment, and also the day we launched a massive script using Artificial Intelligence (AI) to 'synthesize' 60,000 new counseling data points to fill our dataset gaps.

April 5, 2026, marked a decisive operational turning point for the Maeumjigi capstone project. Having completed four iterative versions of data cleansing (v1 through v4), our team needed to accomplish two critical engineering objectives:

1. Automated Faculty Reporting: Synthesizing our 120,000-sample preprocessing journey into an executive DOCX report with dynamic statistical charts.
2. Next.js 14 Frontend Scaffolding: Launching the client-side architecture capable of handling low-latency server-sent events (SSE) for streaming model responses.

This is the record of April 3, the day we began in earnest the task of refining the training conversation data to feed into the deep learning model (Llama 3.1). It was also the day we ran a massive pipeline that makes the Artificial Intelligence (AI) read the text and judge the emotion on its own, going beyond simple data merging.

April 1, 2026, was the day our engineering team launched the automated preprocessing pipeline for the massive counseling corpora that would form the foundation of Maeumjigi. Training a high-performance open-source Large Language Model (LLM) like Meta Llama 3.1 requires rigorous data hygiene: converting unstructured multi-party dialogues into clean, token-bounded, and special-token-aligned instruction datasets.

This is the record of March 31, when we conducted the Android build test of the MND Convention reservation alarm app, which was the initial direction of the Maeumjigi project, while simultaneously starting the Llama 3.1 fine-tuning data preprocessing, hinting at a full transition into an AI-based project.

This is the record of March 30, when we officially started developing the 'MND Convention Reservation Alarm' application, which was the initial idea for the Maeumjigi project. Based on the coding rules we set last time, we established a detailed project plan and laid the foundation for Test-Driven Development (TDD).

March 26, 2026, marks the historic first day of our university capstone engineering project. Today laid the structural bedrock for everything to come: adopting an autonomous AI coding assistant (Claude Code), designing team collaboration protocols, and making the pivotal architectural decision that transformed a simple reservation alert tool into Maeumjigi, a specialized AI mental health counseling platform.