Agents that read the web, tools that clean data nobody wants to clean, models that watch the road. Have a look around — each one started as somebody's annoying afternoon.
Give an LLM a URL and a goal in plain language — it drives a real browser and answers, never seeing a line of HTML. Perception falls back accessibility tree → DOM → vision, and a safety layer hands off before it ever touches a password field.
An AI data-cleaning workbench for financial analysts. Eight transforms, automatic type detection, and an LLM that proposes a fix you approve before it runs. Raw cell values never leave the browser — only a schema summary goes to the model.
Turns a consultation and a stack of medical reports into structured, guideline-driven clinical documentation — so oncologists spend their evening with patients instead of paperwork.
A Telegram bot that turns messy expense chatter and photographed cheques into structured transactions. The client stopped keeping a spreadsheet; their CA gets clean books automatically.
An AI DJ that hears structure: beat, key and cycle analysis, automatic cue-point detection, then renders a continuous beatmatched set. No crossfade guesswork.
A real-world RL environment that grades whether an agent can actually run a support inbox — escalate the outage, answer the routine ticket, spot the spam — across progressive difficulty tiers.
YOLOv8 tracking that works out where the road actually is — inferring road, sky and mid lines so it adapts to any camera angle — then computes time-to-collision and warns on cut-ins.
Reads audit workpapers the way a reviewer does and surfaces what needs a human — an intelligence layer over documentation that used to be checked line by line.
Statement preparation that stops being a manual assembly job — structured extraction and generation for a workflow that ran on copy-paste.
Classical computer vision done properly — perspective transform, colour and gradient thresholding, curvature fitting — for lane finding on real driving footage.
The path I wish existed when I started: a step-by-step route from knowing no machine learning to being genuinely useful at computer vision.
House hunting as a data problem — gathering and structuring listings so the shortlist comes out of evidence instead of endless tabs.
An early-stage product build. Repo's public, story's still being written.
// Nothing under that filter yet — try another.
Hackathon builds, coursework that outgrew its brief, and things I made to answer a question. Not everything needs a case study.