Research and software

Current research

Orchestr.AI

Training, inference, and benchmarking across SchNet, PaiNN, SO3net, FieldSchNet, NequIP, Allegro, and MACE.

One YAML file, SLURM, and multi-GPU.

Current research

Quantum-dot force fields

Machine-learning force fields for quantum dots at BCMaterials.

Data, training, and evaluation pipelines.

Current research

Machine-learning interatomic potential foundation model for quantum dots

One potential for quantum dots, rather than a new force field for each system.

Current research

Drug discovery

Solubility, bioactivity, and pIC50 from molecular structure.

Language models

Research agent

Multi-agent research system with Gemini 1.5 Pro and web search.

Automated research.

Language models

LangGraph assistant

Research assistant that searches the web and synthesizes the result.

Language models

Synthetic data GAN

GAN-generated data for privacy-preserving learning.

Language models

Code generation

Salesforce CodeGen with Hugging Face and the GitHub API.

Language models

Document analysis

Extraction and summary with T5 and RoBERTa.

Language models

Resume screening

Skill-gap notes from PDFs with Llama, Ollama, and PyMuPDF.

Language models

Web-search assistant

Llama, LangChain, and DuckDuckGo.

Language models

Hybrid expert assistant

ChromaDB, BM25, Groq, and a Chainlit interface.

Language models

BERT emotion and sentiment

BERT and EmoRoBERTa for sentiment and emotion.

Language models

SQuAD questions

Question generation and answering, including a multitask setup.

Language models

Streamlit language tools

Summaries, question tools, answer scoring, and YouTube transcripts. text-davinci-003, LangChain, Streamlit.

Language models

Sentence similarity

MiniLM, MPNet, DistilRoBERTa, and spaCy.

Molecules

PRC2 inhibitors

PaDEL descriptors, screening, docking, and molecular dynamics.

Molecules

MPNN pIC50

Message-passing network fused with ChemBERTa, plus a Streamlit QSAR dashboard.

Molecules

Multi-task activity

Mol2Vec embeddings, shared layers, separate heads for activity and pIC50, focal loss, virtual screening.

Molecules

QSAR platform

Curation, RDKit descriptors, fingerprints, PyCaret, and screening in one app.

Molecules

ChemBERTa and ChemGPT

Embeddings from ChemBERTa and ChemGPT, including ChemGPT-1.2B, with a random forest.

Molecules

ChemBERTa RNN

Bioactivity from ChemBERTa embeddings through a recurrent net.

Molecules

Graph attention

pIC50 with a graph attention network.

Molecules

Chemprop

Bioactivity with bond message passing.

Language

Roman Urdu POS tagger

Master’s thesis. Corpora, spelling normalization, a tagset, a phonetic hash, and an LSTM.

COMSATS, CGPA 3.46.

Language

Naive Bayes sentiment paper

Implementation of a published naive Bayes plus POS-tagger sentiment method.

Language

Roman Urdu sentiment

Deep sentiment model for Roman Urdu.

Language

Twitter keywords

Region-specific keywords from PSL tweets.

Language

IMDB sentiment

Sentiment on movie reviews.

Language

POS taggers compared

Decision tree, HMM, a neural net, and an LSTM.

Prediction

OLX prices

Deep model for listing prices.

Prediction

Ozone

PyCaret regression for O3.

Prediction

Dark patterns

NLP classifiers for dark patterns on shopping sites.

Prediction

COVID maps

GeoPandas and Folium, with marker clusters.

Prediction

Phone prices and maps

AutoML prices with seller locations on a Folium map.

Prediction

Marketplace maps

Seller locations and regional prices.

Text mining

COVID abstract clusters

k-means and hierarchical clustering, with intra- and inter-cluster scores.

Text mining

Semantic clusters

Documents grouped by shared meaning.

Text mining

Comment clusters

k-means and text similarity on user comments.

Text mining

COVID topics

LDA with coherence tuning on scientific abstracts.

Text mining

Author topics

LDA and NMF on the Spooky Author set.

Text mining

Keyword extraction

YAKE, RAKE, TextRank, and KeyBERT for product text, reviews, and news. Streamlit comparison.

Text mining

Market basket

Apriori frequent itemsets and a retail recommender using season and region.

Vision and biology

Vegetable images

CNN classifier.

Vision and biology

Chest X-rays

Pneumonia detection with fastai.

Vision and biology

Brain tumors

CNN classifier.

Vision and biology

Sequence alignment

Fully convolutional network.

Vision and biology

Sequence classification

Protein, DNA, and RNA with AdaBoost, forests, and deep models.

From scratch

Weighted ID3

The enhanced ID3 paper, with Graphviz and Turtle visuals.

From scratch

Hidden Markov model

Sequence prediction for assignment types.

From scratch

Apriori

Frequent itemsets on the Zoo dataset.

From scratch

Genetic algorithm

Team-performance search.

Earlier work

API and shop collection

Scrapers for APIs and shopping catalogs, used in agency and Fiverr work.

Earlier work

Prices and jobs

OLX prices, LinkedIn job posts, and CoinMarketCap series.

Earlier work

Windows and forms

PyAutoGUI, form filling, and scheduled posting.

Earlier work

Poetry scrapers

Collected poetry and SMS text from a set of Urdu poetry sites.

Earlier work

Manga chapters

Downloaded every chapter's images and wrote the WordPress HTML.

Earlier work

Video sites

From the download on vidcloud9 and vidstreaming through to a post.