Adjetei Martey

Work

One product in beta, three LLM apps you can try right here, and the computer vision projects behind them.

Featured project

Unstuck

An AI to-do app for people who freeze at a long list.

Live · invite-only beta

You throw your whole messy list at it. It groups the tasks into batches you can do in one go, picks where to start, breaks scary tasks into small steps, and writes the exact words for phone calls and emails, with nothing left for you to fill in.

What I built

  • LLM pipeline that sorts, batches and breaks down tasks, with call and email scripts that tell admin calls apart from personal ones
  • Accounts and cross-device sync on Supabase, with row-level security and a code-based closed beta
  • Reminders that arrive as real push notifications, sent by scheduled jobs
  • Focus mode, voice input, accessibility settings and five languages
  • Next.js
  • TypeScript
  • AI SDK
  • Claude
  • Supabase
  • Web Push
  • Vercel
letsunstuck.com
Unstuck: adding a messy list of tasks, with one marked as tough
Unstuck: tasks sorted into batches, with the first one highlighted as Now

Live demos

FrontDesk: a guest books a table for six with a nut allergy noted, and the booking appears in the owner's view

Booking agent

FrontDesk

Answers your guests' questions and takes bookings day and night, and hands the tricky ones to you.

An AI host for a demo restaurant. It looks up the menu and allergens, checks free tables, books only after the guest confirms, and hands big groups, private hire and complaints to staff. The owner's view shows bookings, hand-offs and every tool call.

  • AI SDK
  • Tool calling
  • Next.js
TriagePilot: the agent trace, with the classify tool rating an API outage as urgent

Tool-calling agent

TriagePilot

Sorts your support inbox and drafts replies for your team to approve.

An agent that triages support tickets step by step: it classifies the ticket, searches a knowledge base, then drafts a reply. Every tool call and result is shown as an inspectable trace.

  • AI SDK
  • Next.js
  • TypeScript
docuchat: an answer about annual leave from a sample handbook, with cited source chunks

RAG

docuchat

Lets your staff get answers from your own documents in seconds, with sources.

Upload PDFs, Word docs or notes and ask questions. Answers come only from your documents, with citations to the exact chunk they came from. Hand-rolled chunking, embeddings and cosine retrieval, no vector database.

  • AI SDK
  • Embeddings
  • Next.js
Extractly: fields and line items extracted from a sample invoice, ready to edit and export

Structured extraction

Extractly

Turns piles of invoices and forms into spreadsheet-ready data, no retyping.

Turns invoices, receipts and CVs (images or PDFs) into clean, schema-validated JSON using a vision model, with inline editing and CSV or JSON export.

  • Vision LLM
  • Zod
  • Next.js

Research & ML

  • Chest X-ray classifier · Computer vision

    DenseNet121 (the CheXNet architecture) fine-tuned on NIH ChestX-ray14 to spot 14 findings at once, trained with BCE loss and evaluated with per-label AUROC.

    • PyTorch
    • DenseNet
    • Multilabel
    GitHub
  • Facial emotion recognition · Computer vision

    Reads seven emotions from face images in FER2013. A VGG19 fine-tuned with class-weighted loss reached 68.6% accuracy, against 52.7% for a custom CNN built from scratch.

    • PyTorch
    • fastai
    • Transfer learning
    GitHub

Hiring, or have a project in mind?
Let's talk.

Tell me what you're trying to do and we'll set up a free call to talk it through, including cost.

adjeteimartey@gmail.com