Hi everyone,
I’m Bladimir, an AI automation engineer based in Bucaramanga, Colombia (UTC−5, no daylight
saving). I’m opening up contract capacity and I’d rather show what I’ve shipped than list tools
I’ve clicked through.
What I actually run in production
I co-founded a small AI automation company and built its main product: an end-to-end
invoice-processing system for Colombian e-invoicing (DIAN). It has been live since July 2026 at a
hardware manufacturer with ~30 years in the market and has processed 220+ real invoices with
no manual data entry.
The pipeline, all orchestrated in n8n (~590 nodes):
- Intake — dedicated email alias, attachment and ZIP handling, XML parsing of the DIAN
invoice format. - AI layer — 8 Claude API calls per invoice, mixing Haiku and Sonnet by task: expense
category, risk level, duplicate detection. - Human in the loop — the approval request is routed to the right department and can be
approved straight from the email with one click, no system account needed. That single
design choice is why the system is actually used: 79% of approvals come from people who don’t
have a login. - Write-back — ERP integration, full audit trail in PostgreSQL.
Around it: multi-tenant Supabase/PostgreSQL with per-client isolation and row-level security, JWT
auth with TOTP, a React dashboard on Vercel, and regression test benches I wrote myself that
run outside n8n and must go green before any change to the parser or the ERP payload ships. That
last part is the one I’d bring to your project: I’ve been burned enough to test before deploying,
not after.
What I can help with
- n8n workflows that have to survive contact with real users — error handling, retries, idempotency,
kill switches, and a rollback story. - Claude API integrations: agent design, prompt engineering, RAG, anti-hallucination layers,
evaluating outputs instead of hoping for the best. - Document and invoice pipelines (XML/PDF → structured data → downstream system).
- Supabase / PostgreSQL data modelling for multi-tenant apps.
- Rescuing workflows that “work on my machine” and fall over in production.
Practical details
- Availability: 30–40 hours/week on contract. Open to a long-term engagement, not just a
one-off build. - Timezone: UTC−5, no DST — full overlap with US business hours, mornings with Europe.
- Languages: Spanish native, English C1 (13 months as a live medical interpreter before this,
so client calls are not a problem). - Rate: US$20–25/hour depending on scope, or a monthly retainer from US$3,000.
- Contact: blado0209@gmail.com · linkedin.com/in/bladimirculman / Portfolio (architecture diagram): Bladimir Culman — Portfolio - Google Drive
Happy to walk through the architecture on a call, or to do a small paid pilot task first if you’d
rather see the work than read about it.
Bladimir