Qasim Shah
All projects

Agent workflows

Fieldwork · Applywell

Research and preparation workspaces where sources, drafts, queues, and user decisions stay together. Every consequential step stays reviewable.

Status
In development
Kind
AI & systems
Stack
TypeScript · FastAPI · SQLite · Agents SDK
Try the machine: gate a review
Fieldwork research pipeline with fictional company fixtures and review controls
Local Fieldwork capture · fictional company fixtures

The problem

A fast assistant can still lose a source, overwrite a user's edits, or mistake a prepared draft for completed work. I wanted the workflow to make those differences explicit.

What was built

  1. Fieldwork: prospect imports, company and contact evidence, deduplication, suppression, notes, outreach drafts, and filtered exports.
  2. Applywell: employer-board imports, an applicant-reviewed profile, persistent preparation queues, draft review, and export.
  3. Version-bound approval states and durable records that distinguish preparation, review, and reported completion.

What I learned

Preserve the human's edits and bind approval to the version they actually reviewed. Cancellation must reach persistence, and a draft needs a different state from a submission.