Hey, I’m
Building intelligent systems
that solve real-world problems.
I work with machine learning, deep learning and generative AI to turn data into useful, reliable products — systems that explain their reasoning instead of asking to be trusted.
Building intelligent systems
that solve real-world problems.
I work with machine learning, deep learning and generative AI to turn data into useful, reliable products — systems that explain their reasoning instead of asking to be trusted.
I’m a Computer Science undergraduate at VIT Chennai specialising in AI and machine learning. I build systems that put the evidence in front of you instead of asking to be trusted — a weld detector you can interrogate, a recovery pipeline graded against ground truth, a matcher whose score you can reproduce by hand.
“PhotoRec tells you it found 9,000 files. This tells you which 40 are your photos, and why.”
“Most tools hand you a score. This one shows its work.”
“A detector that only prints 'defect: 0.81' asks you to trust it. This one is built the other way round.”
Classical models trained, cross-validated and calibrated until the reported confidence matches what actually happens. Seven classifiers benchmarked in one command; the best one persisted automatically.
Detection pipelines owned end to end — dataset curation, labelling, GPU training, then the honest part: diagnosing the gap between a 99% training score and a 74% held-out one.
Multi-agent graphs where each stage owns one responsibility, retrieval that cites what it found, and language models kept strictly downstream of the number they are describing.
The part that makes a model a product: typed APIs, Docker images that carry their own weights, CI across a Python version matrix, and tests that fail when a claim stops being true.
Idle cloud resources sit unused and forgotten, quietly driving up cost and widening the attack surface. This work presents an event-driven Boto3 system that scans EC2, EBS, S3, IAM and RDS for waste, then recla…
Hardening both for real use — baking model weights into the Docker image so a cold start can’t fail on a fetch, and pushing the recovery pipeline’s test coverage into the statistical checks that decide a verdict.
Multi-agent graphs are easy to demo and hard to keep honest. I’m working through state, retries and partial failure — what an agent does when the stage before it returned something almost right.
Following the thread from MediXplain and the XAI drift work: how far can an explanation be pushed before it stops describing the model, and can the confidence number be trusted at the same time.
Still carrying a camera. Concert lighting is the hardest exposure problem I know of that has nothing to do with code, which is exactly why it’s a good break from it.
