Veriflow
An agent platform that has to prove every claim it makes before a human signs off.
42 projects shown
An agent platform that has to prove every claim it makes before a human signs off.
Hybrid rule-based and LLM-assisted mapping for B2B EDI documents.
A RAG chatbot that knows when to look something up exactly instead of searching for it.
A Transformer trained from scratch on a 24,525-pair parallel corpus.
Speech, frames, and a language model, run in parallel to summarise a video.

What transfer learning buys on CIFAR-10: 69% from scratch, 97% fine-tuned.
Live demoPaint a few strokes and let colour clusters separate foreground from background.
Live demoA from-scratch nearest-neighbour classifier on 32 x 32 faces, compared with SVM and Naive Bayes.
Live demoThe original learning rule, implemented from scratch and animated.
Live demoThe operation behind CNNs, written by hand and compared with SciPy.
Live demoSEND + MORE = MONEY, solved by search, with the frontier on screen.
Live demoLamport's mutual exclusion algorithm, one memory operation at a time.
Live demoCars, buses, a mutex, and two condition variables.
Live demoRegression baselines, stationarity tests, and walk-forward evaluation on daily BTC prices.
Live demoRule-based sentence boundaries for Urdu, where punctuation alone is not enough.

Open-vocabulary detection turns a photo of produce into a bill.

A compact Keras CNN, augmentation, and a learning-rate ablation.
Random search over nine hyperparameters for an MNIST CNN and an IMDB BiLSTM.
Eight forecasting models behind one Streamlit interface.
A HYBRIDJOIN stream-relation join loading a star schema, then OLAP in SQL.
From nested scraped JSON to a normalised SQLite database.
What is actually inside an Urdu speech dataset.
MFCC fingerprints and MinHash banding to find similar tracks.
Normalised engagement metrics and median baselines for fairer content comparison.
Smaller assignments and experiments, kept for completeness. Each links to its repository.
A convolution plus mixture-of-experts acoustic model prototype and a Whisper baseline for Urdu.
PyTorch · torchaudio · CTC · Whisper
Next-word prediction on Tiny Shakespeare with random versus Word2Vec embeddings.
PyTorch · gensim
Five-class sentiment from meme text with TF-IDF classifiers.
scikit-learn · NLTK
Bag-of-words classifiers with requirements and design documents.
scikit-learn · NLTK · XGBoost
Cleaning, stemming, lemmatisation, n-grams, and WordPiece on news text.
NLTK · Transformers
Stroke-seeded colour clustering (the Lazy Snapping Q1 notebook).
scikit-learn
Imputation, encoding, and derived features on the Titanic dataset.
pandas · scikit-learn
A Jira-like relational schema with a role-based Express front end.
MySQL · Express · EJS
Simulated sales events streamed through a Kafka topic.
Kafka · kafka-python
Express and MongoDB REST API with JWT auth and calendar sync.
Express · MongoDB · JWT · Socket.io
PGM statistics, negatives, and run-length coding with hand-built data structures.
C++
Layers as processes, neurons as threads, pipes between them.
C++ · pthreads · fork · pipes
Projects whose code is not public. Described from the resume only.
Chunking, embeddings, vector search, reranking, conversation memory, and query rewriting for grounded multi-turn question answering.
A tool-calling agent for PDF parsing, semantic search, structured summarization, and retrieval-augmented question answering.
CLIP-based zero-shot recognition with transformer captioning, plus CNN face-mask and emotion models optimized for low-latency webcam inference.