Projects | Aditya Rajesh

A forecasting project for reasoning about natural gas storage levels and market-relevant inventory changes.

Problem: Energy market data is noisy, seasonal, and sensitive to weather, demand, supply, and calendar effects. A useful forecast needs to make those assumptions visible instead of hiding them behind a single number.
What I built: I am building a reproducible forecasting workflow that frames storage changes as a time-series problem, compares baseline and machine-learning approaches, and keeps evaluation explicit.
Stack: Python, pandas, scikit-learn, statsmodels, time-series validation
What I learned: Forecasting quality depends as much on feature design, leakage control, and evaluation windows as it does on model choice.
Status: in progress
Links: Not public yet

A practical AI-assisted workflow for reading research papers more carefully and consistently.

Problem: Research papers are dense, and it is easy to summarize them without checking study design, assumptions, evidence quality, and limitations.
What I built: I am prototyping a structured assistant that helps extract claims, identify methodology, flag limitations, and produce a critical appraisal rather than a generic summary.
Stack: Python, LLM prompting, structured outputs, document parsing
What I learned: The useful part of AI assistance is not speed alone; it is forcing a repeatable review process with explicit questions and failure checks.
Status: in progress
Links: Not public yet

A small language-modeling project for understanding token prediction from first principles.

Problem: Modern language models are difficult to reason about if the basics of sequence modeling, loss, sampling, and training loops stay abstract.
What I built: I built a character-level model starting with bigram probabilities and used it to understand how simple statistical structure turns into generated text.
Stack: Python, PyTorch, probability, model evaluation
What I learned: Even simple models make the core workflow concrete: represent data, define a loss, train against examples, sample outputs, and inspect failure modes.
Status: complete
Links: Not public yet

A minimal Astro site for writing, projects, and public notes.

Problem: A portfolio needs to be more than a list of links; it should explain what I work on, how I think, and where the work is going.
What I built: I rebuilt the site as a static Astro project with a small content system, clean navigation, dark/light mode, and focused pages for projects, writing, and current work.
Stack: Astro, TypeScript, CSS, GitHub Pages
What I learned: A simple site is easier to maintain when the structure is clear and the content is specific.
Status: in progress
Links: GitHub