AI engineering
Model workflows that stay grounded, bounded and reviewable.
Agentic bid writing and assessment graphs at work; a five-agent writing pipeline, structured document edits and a 30-prompt engine in my own products.
Capability atlas / project-linked
No logo cloud and no invented percentages. Every capability is attached to a product where it shaped a real architecture, interface, safety boundary or deployment.
Model workflows that stay grounded, bounded and reviewable.
Agentic bid writing and assessment graphs at work; a five-agent writing pipeline, structured document edits and a 30-prompt engine in my own products.
Picking the provider per job and keeping a way out when it fails.
Custom clients with fence stripping, JSON validation, retries on 408/429/5xx and labelled deterministic fallbacks.
Checks you can explain, before a model grades a model.
Rule-based audit with three refinement rounds, five hard gates before any AI score, and one-time approval bound to a SHA-256 form snapshot.
Services, pipelines and the contracts between them.
FastAPI AI services and recoverable ingestion at work; a 51-source feed pipeline and a 17-command Python CLI on my own.
Interfaces for complex work: editors, dashboards, extensions and devices.
B2B workspaces, rich editors, a canvas wallpaper studio, an offline mobile app and this desktop OS.
Software that lives inside other software.
A Chrome extension with verified writes across frames and five ATS adapters, and a Marketplace theme suite with 300+ installs.
Identity and tenancy designed in, not bolted on.
Shared identity migration with claim-first checks and fallback, safe public projections, private storage and hashed OTPs.
Shared state and live media, with their limits stated.
A custom Yjs transport over Supabase and a native WebRTC mesh for research rooms.
Getting it shipped and keeping it green.
Unit suites and CI across products, Playwright smoke checks and video capture, and container delivery on Azure and AWS.