Career Companion App: Wees 3 and 4 Monday Standup

Date: May 26, 2025

Still Learning and Pushing Through

So far, developing the Career Companion app has been a crash course in cloud services, language modeling, and debugging that doesn’t always go as planned. I’m building the app to help users improve their resumes, identify their skill gaps, and get personalized career guidance. At its core, it’s about making job-seeking less overwhelming and more strategic, using AWS and NLP.

What’s Done So Far:

I started with setting up the AWS environment, including S3 for storing resume data, and Lambda functions for handling input and triggering analysis.

I used Amazon Comprehend and a custom token classification model to identify skill entities in resumes. For this, I trained a BERT-based model using Hugging Face's Trainer on annotated data in IOB format.

My training script is now running smoothly. At first, I hit a bug where I mistakenly tried to map label IDs twice, which led to a KeyError. Once I realized I was passing numbers where strings were expected, I fixed the mapping logic.

I also had to rework the token-label alignment. The dataset was initially a single stream of tokens, which made splitting into train and test sets impossible. Segmenting by sentence (based on blank lines in the IOB file) fixed that.

Most recently, I removed an unsupported parameter (evaluation_strategy) from the training arguments to prevent a crash during model training. It was a small mistake, but it took time to trace because the error message wasn’t immediately clear about what was going wrong.

What’s Left:

I still need to wrap the trained model in an inference-ready Lambda function, then deploy it via an API Gateway so that users can submit resumes and get insights back.

I’ll also be adding a web interface using Amplify. That way, the app is accessible even to people without any technical background. The UI will include resume upload, skill feedback, and maybe even suggested job roles using pre-trained embeddings or OpenSearch.

After that, I want to fine-tune the recommendation engine based on skill gaps—matching users to free resources or certifications to bridge the gaps.

I’ll also need to monitor costs and usage on AWS. Since this is a personal project, I have to be careful about services that can get expensive when scaled.

This whole experience has taught me more about deployment pipelines, model fine-tuning, and AWS than any tutorial could. If you want to follow along or test the app when it’s up, let me know. I’m still learning, but it’s starting to come together.