Senior Python Engineer (AI Agents & Browser Automation) - FT/PT - EMEA/LATAM/NA

The company tests whether AI agents can complete real tasks on a business’s website, then fixes what blocks them and keeps monitoring. Three services: Agent Readiness (local service businesses), Aisle Agent (online store purchase journeys), Knowledge Readiness (company information prepared for AI agents).

One Python tool runs browser agents (browser-use with Claude) in Docker on a DigitalOcean droplet, with an independent judge, an evidence-backed Hard Finding rule, a cost ledger, a job queue, an operator dashboard and report generation. It is a working prototype with no independent engineering review yet.

WHAT YOU WILL DO

• Review the whole codebase independently; put it in git with a private remote, .gitignore and CI (lint, offline tests, protected-file hash check).
• Fix the ranked defects: the prepare-only guard rejecting the start URL, with an end-to-end offline test (real guard, scripted fake model, local range); the cost ledger overcharging invalid replies; malformed model replies.
• Harden production: network-level blocking of non-GET requests, client-domain allowlist, dashboard login and CSRF protection on a production server, SQLite WAL and off-droplet backups, secrets, telemetry settings, auth-log review.
• Run calibration against the test range, reconcile cost against the provider console, then deploy with a rollback plan and smoke test.
• Design the engine so that new niches and markets (US and UAE at launch) are added as configuration packs and market profiles, not code: a pack format with a schema, a lint check, a scaffold command and a calibration set, with the Med Spa pack as the first built in that format.
• Then build the full tool in the background, in this order: run version metadata, report approval state and an evidence-method field; group reporting and the monitoring scheduler; a Med Spa niche pack; Aisle Agent (payment guard first, then purchase-path policy, video, verifiers); Knowledge Readiness (approved-answer store, findings scan, replay tester). Each automated piece replaces a manual step only after it passes its gate and agrees with a parallel run., * Protected files (audits/, config/, range/) change only with explicit approval and a refreshed hash manifest. No real model call or real-site request without approval and a spend cap.
• One pull request per ticket with raw test output. Local-service behavior must not change when other services are added.
• You get a non-root droplet account, a separate low-limit API key and scoped access; never owner credentials. Weekly demo; NDA before access.

HOW THE ENGAGEMENT IS SHAPED

• Phase 1, about two weeks: a hardening sprint on the existing engine, because the business launches by serving first clients by hand with it. Phase 2: the full build, ordered by what removes the most manual work. Duration of Phase 2 is a joint estimate after your first-week review; the founder’s own estimate is several months.
• You work alongside a non-technical founder who runs delivery and operations; the engine must stay usable by them throughout. Weekly demo and a short Friday written status.

FIRST TWO WEEKS

Days 1 to 2: repository and droplet review, git and CI, written findings. Days 3 to 5: guard fix with end-to-end test, ledger fix. Days 6 to 10: production restrictions, dashboard hardening, backups, calibration run.

• 6+ years of production Python; SQLite or Postgres; Docker and Linux server hardening; GitHub Actions or similar CI.
• Browser automation depth: Playwright or CDP, request interception, redirects, bot-protection behavior; browser-use experience is a strong plus.
• LLM API integration: tool use and structured output, schema limits, prompt caching, token and cost accounting. Anthropic API experience preferred.
• A safety mindset for automation that touches third-party sites; careful secrets and credential handling.
• Disciplined testing: offline tests with fakes, contract tests, tests that exercise the real path.
• Clear written communication; works from tickets with acceptance criteria and reports evidence, not impressions., * Flask, python-docx or reportlab; Shopify storefront and checkout behavior; Google API verification, Zendesk or Intercom APIs; data privacy practice (isolation, deletion, access logs); evaluation and grading of LLM output.