Find the real bottleneck
I start with work people already repeat: reading documents, entering product data, translating content, or navigating complex product data.
I build the system around the model.
Useful AI features need more than a prompt. They need clear permissions, structured outputs, review states, retries, observability, cost boundaries, and a deterministic path when the model is uncertain.
Keep people in control
Models propose fields, values, copy, or actions. The product makes confidence, review, and correction part of the normal workflow.
Engineer the failure path
Permission checks, timeouts, retries, stored state, and deterministic validation keep uncertain output from becoming silent product failure.
Document intelligence
From paper to structured data.
In Papereg, AI is embedded inside a complete document workflow rather than presented as a chat box.
Read the Papereg case studyPDF structure parsing
Uploaded forms are analyzed for labels, sections, field types, and layout so users can review a proposed digital form instead of rebuilding it by hand.
OCR and batch extraction
Photos, scans, handwriting, checkboxes, and multi-page PDFs can become draft records. Failed files can be retried without losing successful results.
Permission-aware assistance
The workspace assistant can answer questions and help create or fill drafts within a user’s permissions. Destructive, approval, and administration actions stay outside the assistant’s authority.
Multimodal commerce
Turn product photos into useful inventory.
ResellerIO combines vision, structured product data, image workflows, pricing context, and channel-specific copy.
Read the ResellerIO case studyExtraction before generation
Photos become a structured draft with brand, category, size, color, material, and condition. Structured data keeps downstream features consistent.
Image preparation
Background cleanup, crop preparation, and lifestyle image generation help sellers create a reusable asset set from the same intake.
Drafts, not automatic publishing
Marketplace-specific titles, descriptions, and tags are presented for review. The seller remains responsible for what goes live.
Project intelligence
Keep engineering context connected.
HermIIS links product planning, decisions, documents, GitHub activity, and delivery analytics so AI can work from the same evidence as the team.
Read the HermIIS case studyGraph-backed context
Tasks, features, decisions, documents, pull requests, and commits remain connected instead of becoming isolated pages.
Evidence-aware assistance
Semantic search, graph expansion, and citation chains help people inspect the sources behind generated answers, reviews, and task proposals.
Suggestions remain suggestions
AI can surface missing work or propose a decision record, while the team retains authority over scope, priorities, and architectural direction.
Integration work
AI inside existing products.
I also integrate AI into mature systems where security, data ownership, and operational safety matter more than novelty.
OAuth-secured product access
Designed and built a Model Context Protocol server that lets ChatGPT integrate with product capabilities through explicit authentication and authorization.
Localization automation
Combined OpenAI and DeepL APIs with an existing internationalization workflow to accelerate translation across Spanish, Chinese, Japanese, and Portuguese.
Agent-assisted engineering
Use coding agents for focused implementation, audits, migrations, tests, documentation, and repetitive maintenance—while preserving review and verification.
“The model can be uncertain. The product experience cannot be careless.”
Build with intent
Have an AI idea that needs product discipline?
I can help turn an interesting capability into a secure, understandable, and maintainable workflow.