Dot matrix with data pipelines
Back

Problems with my resume

· 3 min read

I had a friend have their partner look at my resume and tell me what they think. What I heard actually hit me hard. First glaring issue he found was the lack of numbers in what were my biggest projects - OpenCodeReasoning and Passtor.

OpenCodeReasoning is a project where I had to use the OpenCodeReasoning2 dataset by NVidia to train a model to be able to find which part of a buggy code is problematic, provide textual explanations, and generate a fixed version. The problem is - the dataset itself is LLM-generated. I trained my model using SFT, then with reinforcement learning using GRPO for code correctness. But, GRPO is quite expensive and I ran out of GPU runtime in Colab.

But even then, when I was able to achieve at least a portion of what was possible, there are no numbers to report. I can only say the number of correct codes that the model generated for a handful of coding questions. Not for a chunk of the dataset. Besides the dataset itself is not reliable because it is LLM-generated.

Passtor is a data pipeline that takes in accessible, tagged, annotated PDF files, and creates a dataset of image-caption pairs. I wanted to try creating a step further from what CodeMantra did. These images could be actual figures, or vectors, or even PDF drawing objects. There was a lot of effort put into this and making a whole pipeline all on AWS beginning to end. Used a landing page on S3, presigned url for file upload, S3 to store raw PDF, Lambda to extract these image-caption pairs, DynamoDB to store data about images, Cognito to manage user profiles …. [whew!] … SNS and SQS to notify and update status when job is done…

But once the pipeline was ready, I faced a problem. The repository I initially had known for annotated PDF files is no longer live. The only other option is to ask CodeMantra for their confidential documents *blep*. Futility… Now I cannot measure any latency or throughput, nothing.

So in the end, I did not have numbers to report even though I put so much effort into it all.

Secondly, I don’t have a RAG project. RAG is such a simple concept, yes, and I did do a very simple RAG system once for the fun of it, but that was all. Perhaps now I should try an actual project to add to my resume? Many jobs these days do include RAG after all.

Comments

Loading comments…

Leave a comment

Comments appear after review. Your email is never published.

← All posts