I’m Ansh Agrawal, a second-year B.Tech AI & ML student at Universal AI University. I build machine-learning, deep-learning and LLM applications — a graph neural network, calibrated classifiers, an agent pipeline and a retrieval system — and evaluate each one on held-out data before describing it.
Background: 200 papers from the Cora citation graph, the data GraphMind classifies.
I work across the whole of an applied ML project: preparing data, training or orchestrating a model, serving it behind a small API, and building an interface someone else can actually use.
The step I care about most is evaluation — scoring the saved model rather than trusting a notebook printout, saying where every number comes from, and writing down what the system gets wrong. My current interests are graph learning, agentic LLM workflows and retrieval-augmented generation.
Every card links to a live demo or a recorded run and a full case study. Numbers carry a label showing where they come from — see how I evaluate. Source code is available on request. Live demos are hosted on Hugging Face Spaces.
A two-layer graph convolutional network assigns each of 2,708 papers in the Cora citation graph to one of seven topics, combining a paper’s own word features with those of the papers it cites or is cited by.
A search agent, a reader agent, a writer chain and a critic chain turn a topic into a structured, sourced report with a self-critique. The public demo is rate-limited and its web scraper is protected against server-side request forgery.
A bidirectional GRU assigns one of six labels — sadness, joy, love, anger, fear or surprise — to a sentence and returns the full probability distribution, so uncertain predictions are visible.
A tuned XGBoost pipeline wrapped in sigmoid calibration estimates the probability that a loan defaults, and shows the default rate actually observed among held-out applicants with similar scores. A research demonstration, not a lending tool.
Extracts a video’s audio, transcribes it with Whisper, summarises it with Mistral, and answers questions from the four most relevant transcript chunks — showing that evidence next to every answer.
Screen recordings of real runs, each with a title and end card stating what was measured. They have no sound; a video loads only when you press play.
Classifies paper 1708 and its two-hop neighbourhood, opens a misclassified test paper, runs the edge experiment, and scrolls through the evaluation computed by the server.
One real research run on “CRISPR gene editing” in real time: the four stages, the report, the critic’s score, the sources, and the measured tokens and cost.
Four sentences with their full probability distributions — including a real error (a happy sentence labelled fear) and a factual sentence forced into an emotion — then the test-split evaluation.
A typical application (3.5 %, low risk), a riskier one (94.8 %), an impossible input being rejected, the sensitivity tool, and the held-out evaluation.
An 88-second synthetic meeting is transcribed, summarised and indexed; three questions are answered with retrieved evidence, including one about something never discussed.
A number is only useful if a reader can tell where it came from and what it leaves out.
Metrics are recomputed from the saved model files on the original held-out split, not copied from a single notebook run.
Every figure is marked as re-evaluated, measured in a live run, or copied from notebook output.
A baseline or an ablation where one exists — a random forest without graph structure, the GCN without citation edges, logistic regression for credit risk.
Data leakage, a checkpointing bug, a threshold chosen on the test set, and a speaker-attribution error are documented, including the ones not yet fixed.
Grouped by area, each linked to the project that demonstrates it.
Used in CreditCompass
Used in Verbalyx
Used in all five projects
Used in all five projects
Used in Scholaryx, CreditCompass
Built a university-focused AI chatbot using Python, retrieval-augmented generation and LLM APIs to help users find university information.
Prepared the event documentation for two AI Cell training events: “AI Tools for Hackathon” and “Build Your Own Calling AI Agent Tool”.
Collected community feedback and information through online forms in Kushiwali and nearby villages to support data collection on local needs.
Team Regnum Carya; my role was R&D. We proposed an explainable-AI design for diabetic-retinopathy screening — a design and pitch; the software is not yet built.
in progress
Second year.
I’m looking for AI engineering and research internships. Source code for the five projects is available on request.