A bit about how I got here.
I like the layer between an idea and a working system: APIs, models, data, evaluation and the engineering glue that makes all of it useful.
I'm a B.Tech Computer Science student at K.R. Mangalam University in Gurugram. My current direction is AI backend engineering, with a focus on LLM systems, agentic workflows, NLP and backend infrastructure.
At Ethara AI, I worked on LLM post-training projects involving SFT, RLHF and RLVR, adversarial prompting, evaluation datasets, agentic workflows and multimodal model assessment.
Before that, at Unified Mentor, I worked across full-stack web applications, REST APIs, CRUD systems and relational data using Python, Flask, MySQL and Git.
Outside the stack: badminton, chess, gaming, music and the occasional trip when the schedule permits.
Model behavior, not just model output
At Ethara AI, most of my day was spent probing where LLMs break - adversarial prompts, edge cases, reasoning failures - then turning that into structured evaluation data. I think about AI work in terms of failure modes first.
APIs as the real interface
The UI is rarely the interesting part. I care about the contract underneath it - clean REST endpoints, sane schemas, predictable errors - the layer that makes a system trustworthy to build on top of.
Retrieval and memory over raw context
For anything knowledge-heavy, I default to designing the retrieval layer first - chunking strategy, vector store choice, ranking - rather than throwing everything at a longer context window.
Ship small, verify constantly
Every project I list actually runs - deployed, tested, or at minimum reproducible from a clean clone. I'd rather have three working systems than ten half-finished ones.
The condensed version - everything else lives on the résumé.