What I Build
I don't just advise. I build and run AI systems in production. These are my proof points: the working systems that keep my advice tied to what happens in practice.
shilpiworks
Human imagination. Machine precision.
shilpiworks is an AI-powered creative commerce platform where 16+ autonomous agents continuously generate, validate, price, and publish wisdom stickers, bookmarks, and laser-cut art. The entire pipeline — from theme selection through image generation, OCR validation, and smart pricing to auto-publish — runs on production cron schedules without human intervention.
Built on Next.js, Prisma, Vercel Postgres, and the Mastra agent framework, shilpiworks is a living laboratory for the same AI agent patterns I advise clients on: multi-model orchestration, decision-trace capture, outcome linking, and autonomous operations at scale.
The stickers themselves are “wisdom stickers for readers, leaders, and lifelong learners” — drawing from philosophy, science, literature, indigenous knowledge, and culture. Thirteen curated collections span themes from Stoic philosophy to curious kids to Minnesota wildlife.
What Makes It Work
The technical architecture behind shilpiworks reflects the same patterns I help clients design and deploy.
16+ Autonomous Agents
Themed AI agents (Stoic philosophy, Indigenous wisdom, Science, Women of Wisdom, and more) running on production cron schedules — zero human designers.
Multi-Model AI Pipeline
Gemini 2.0 Flash for speed and cost, GPT-Image-1 for generation, with intelligent fallback and cost optimization across models.
Decision Trace Capture
Every agent run logs its full reasoning chain — theme candidates, style selection, OCR attempts, quality scores. Institutional memory, not black boxes.
Outcome Linking
Agent decisions are connected to product performance — page views, cart adds, purchases, and revenue. The system learns what works.
OCR Validation Pipeline
Text legibility is verified before publishing. Failed stickers are auto-retried with corrections — no broken products reach customers.
Perceptual Hashing
Near-duplicate detection prevents the agent network from repeating itself. Theme diversity is tracked and enforced across all agents.
The same patterns, applied to your business
The architecture, experiment design, and operating habits I use in my own products are what I help clients build. What I learn from running them goes straight into my advice.