thanks for showing interest in my work !!!
my name is Archisa Bhattacharya, i am an aspiring AI engineer studying at Stevens Institute of Technology!
reach out to me at archisa28@gmail.com! i love to chat
choose your move:
rounds played: 31
record: 8W · 18L · 5T
move counts:
🪨 Rock: 12 · 📄 Paper: 10 · ✂️ Scissors: 8
last player: @archipelagoing
last round: 😄 ✌🏼 vs 🫱🏻💻 → 🏆 Win
🪨📄✂️( this runs via GitHub Actions! here are some instructions ) 🪨📄✂️
(1.click your move 🪨 2.click commit 📄 3.wait till round complete ✂️ 4. come back to my profile 🪨 5.refresh the page to see results)
an ambient planning surface for the home
Planoramic turns Fire TV into a shared household calendar that is designed to be understood from across the room.
i'm currently exploring how Fire TV, Alexa+, Amazon Bedrock, Google Calendar, voice, and shared state can work together as one multimodal experience.
voice for intent · TV for understanding · remote for exploration · phone for configuration
studying how relational and situation-sensitive information changes across transformer depth
SituatiONION is an interpretability project centered on GPT-2 XL. It uses controlled paraphrase/counterfactual experiments, layerwise probing, geometric analysis, and representation comparisons to investigate what information is accessible at different transformer layers.
current questions include:
- how paraphrases and counterfactuals separate across depth
- where relational information becomes decodable
- how different token/readout choices change what we observe
- whether apparent mid-layer structure survives stronger controls
- how representations generalize to unfamiliar compositions
☁️ PCA Cloud
turn your spreadsheet into clouds
an interactive browser-based tool for exploring high-dimensional datasets using PCA and t-SNE.
upload a CSV and explore its projected structure as an interactive 3D cloud, compare dimensionality-reduction methods, inspect explained variance, and save visualizations.
game theory + evolutionary dynamics for multi-task learning
an experimental extension of Nash-MTL that combines:
- Nash bargaining for stepwise task fairness
- replicator-style dynamics for longer-term task adaptation
- gradient-based multi-task optimization
the project asks whether evolutionary scheduling can help prevent tasks from being persistently neglected during shared-model training.
what if a decision-making system behaved more like a coalition than a single agent?
a small experimental cognitive architecture where multiple internal agents compete and cooperate over a shared concept graph.
agents representing reactive, habitual, deliberative, and social strategies propose actions, while a coalition policy determines the system's final behavior.
🧴 Tenue
a social identity layer built around fragrance ownership and presence
Tenue explores how physical fragrance ownership, wear history, and communities around individual fragrances could form a persistent social network.
built around:
- ownership
- presence
- fragrance communities
- cultural activity
- personal wear history
there's a lot more scattered around here: NLP experiments, data visualization, cybersecurity labs, Chrome extensions, statistical tools, and assorted ideas that became repositories.
some are polished projects. some are experiments. some are evidence that i should probably stop starting new repositories.
that's part of the fun :-)