Killed by Tin 🪦
Not all projects make it to production. Some die due to complexity, some get replaced, and some just vanish. Here lies the resting place for those fallen initiatives.
R.I.P
In-House Knowledge Hub Pilot
2025
A pilot effort of RAG as internal knowledge hub. Gone, but caught a critical design bug for the team.
R.I.P
Internal OpenAI Playground
2024
Playground for piloting ChatGPT internally. Still occasionally used by rebels to test latest GPT-xxx models.
R.I.P
'One-Click' Training UI
2024
A Simple Training UI on HuggingFace Space for non-technical members to finetune models at ease. Company pivoted to using third-party APIs more than finetuning.
R.I.P
Training Runner
2023-2024
Agnostic runner service to kick off training jobs identically everywhere (local, server, SageMaker, AWS Batch, ...). Killed when the company pivoted entirely to third-party APIs.
R.I.P
EMS: Experiment Management System
2021 - 2024
A unified experiment tracking system designed to improve reproducibility across all Cinnamon models. Abandoned when the company shifted focus away from DL research.
R.I.P
Kustomizable Flax
2021 - 2022
'All-in-one' Auto-Training and Pipeline Customization Platform. Way too complicated to adapt into existing ecosystem. Crushed under the weight of its own complexity.
R.I.P
kf-sdk
2021 - 2022
A standardized SDK injected into AI libraries to provide seamless Kubeflow integration and Cloud Storage support. Still used by some libraries to interact with S3/GCS.
R.I.P
kf-playground
2022
Interactive tutorials and guidelines customized for Cinnamon models, created to soften the Kubeflow learning curve for AI Researchers. Scrapped when the company ultimately dropped Kubeflow.
R.I.P
CURE Framework
2019 - 2021
Rigid standardized training framework for all models. Collapsed under massive maintenance debt after the main author left.
R.I.P
OCR Correction Research
2020
Fusing visual and seq2seq context features with radical output for OCR correction. Dead due to long inference times outweighing minor accuracy gains.
R.I.P
ARKnet OCR Model
2019
Multi-head CRNN models using Kanji decomposition. Abandoned and forgotten after pivoting to more pressing project tasks.