Shipping Backend Flows at profound.me

At profound.me, Karan built the hiring and candidate flows and pushed them to the production application in Python. The work took approximately three months from initial requirement to production.

The team adopted a structured approach to agent-assisted development: documented coding standards, architecture guidelines, and a testing policy covering automated, integration, and segregation tests. Karan applied all of these in his work. The standards allowed the team to move faster because known patterns and automated checks caught issues that would otherwise require manual review, reducing the iteration cost of agent-generated code.

Karan also built a dedicated evaluation pipeline for the voice-to-voice product. The pipeline is built using PyCATS and a combination of open-source and commercial models for ASR and TTS, and allows users to communicate with the Profound AI app to share their professional journey.

After a user report flagged ASR misspellings, Karan confirmed the issue through the evaluation pipeline. He gathered the misheard words and a set of phonetically similar terms, then fine-tuned the model on those cases for correct spelling and pronunciation. The process was iterative: he ran evaluations and audits multiple times. By his recollection, the words surfaced in user reports were fully resolved, and roughly 95% of the additional cases found through evaluation came back correctly, with around 5% still incorrect.