Fifteen projects, each one written up properly.
Some of these paid the bills; some existed only because I wanted to know how the thing worked. Every one has a page explaining the problem, the architecture, the parts that broke, and what I would do differently. Filter by what you care about — or hit / and search.
Works out which vLLM settings your GPU can actually run. Type a model id, pick your card, and get a working vllm serve command, a vendor-correct Docker setup and an honest VRAM verdict — including how many concurrent requests will fit. Runs entirely in the browser, nothing to install.
A reproducible data-cleaning engine for Python. An LLM writes the cleaning plan once; that plan compiles to a YAML recipe which replays in pure pandas with zero API calls, tracks every changed cell, and refuses to run when the incoming schema has drifted.
Multi-tenant SaaS that turns insurance loss-run PDFs and scans into structured claims data, pairing a custom OCR engine with domain-tuned LLMs. Schema-per-tenant isolation in PostgreSQL, 92% extraction accuracy, and under two minutes per document.
Meeting recordings in, structured minutes out. FFmpeg normalizes the audio, a Celery worker routes it to one of six speech-to-text backends — including Whisper on your own machine, no API key — and an LLM writes the summary, decisions and action items. Both layers swap at runtime.
A review desk for ASR output. FFmpeg converts any format to 16 kHz mono WAV, splits it into 30-second segments, and Whisper transcribes them in daemon threads. Type-based job locking and JSON-persisted progress, so no Celery and no Redis.
Semantic search over your own files. PDF, DOCX and TXT get chunked with overlap, embedded with DistilBERT and indexed in FAISS. Chunk size, overlap, model and top-k are all knobs, because retrieval quality is an empirical question.
Content-based movie recommendations at scale. TF-IDF vectors over genres, keywords, companies and plot text, compressed from 20K features to 500 dimensions with SVD. Sparse storage turns a 40GB dense matrix into 150MB; queries land under 50ms.
Seven-class skin lesion classifier trained from scratch on HAM10000. The custom CNN head uses 1x1 and kernel-sized convolutions instead of dense layers, with a concatenation-skip ResNet variant alongside. Ships with a Flask API and a webcam service.
An interactive Python course where every lesson is a hand-authored memory snapshot — frames, bindings, heap objects, refcounts — that you step through like a debugger. Five sessions and around 100 steps, running as static files.
FC 25 player similarity over 34 min-max normalized attributes. The full pairwise cosine matrix is precomputed at training time and pickled, so Flask loads it once at startup and every query becomes a row lookup plus a couple of filters.
Three autoencoders — PCA-style linear, convolutional and denoising — sharing one encoder/decoder interface behind a five-mode CLI. A 32-number code gives 96x compression on STL-10 and drives k-NN image retrieval and latent-space morphing.
Regression and classification in raw NumPy: the normal equation, its ridge form with an unpenalised intercept, and logistic regression by gradient descent on binary cross-entropy. Every algorithm derives the maths before it writes a line of code.
This site. Static HTML pages over one CSS custom-property design system, with pointer-tracked card glows and Intersection Observer reveals in plain JavaScript. No framework, no bundler, no build step — pushing to main is the whole deploy.