Portfolio

Problem: For elderly storytellers, installing an app or signing in is the barrier itself, and raw audio and transcripts are the most sensitive data the product holds.

Approach: An AI voice agent places and answers calls on a regular 070 number; consent comes from a verification code texted to the storyteller's own phone. Audio and transcripts sit behind Supabase RLS and private Storage, visible only to the memoir's owner and the operator, and audio is deleted after 90 days.

Result: Stories arrive through three paths — phone, web voice chat, and uploaded recordings — with family invites and link sharing; tests pin that all 25 admin-only database functions reject ordinary members.

Problem: Coordinate useful work across a multi-repository ecosystem without lowering review quality.

Approach: Built an issue-to-implementation workflow with isolated branches, automated checks, review gates, and maintainership feedback loops.

Result: 789 pipeline-merged pull requests within 881 organization-wide merges; the live network at agentgram.co has 520+ registered agents and 3,600+ posts; agentgram has 32 GitHub stars.

Problem: Keep long-running autonomous work observable, recoverable, and aligned across specialist agents.

Approach: Combined scheduled jobs, shared kanban state, role boundaries, heartbeat checks, and evidence-gated completion.

Result: 45 scheduled jobs across the fleet and 383 completed kanban cards.

Problem: Inconsistent code quality across 5 backend engineers and a constrained infrastructure budget.

Approach: Documented standards, required pull-request review, used GCP credits, and codified infrastructure with Terraform.

Result: Code-review frequency increased 400% (about 5×); recovery time fell from 56h to 5h.

Problem: Parse mathematical notation while rendering editable diagrams and graphs in one workflow.

Approach: Combined MathJSON and Compute Engine with Konva and JSXGraph in an accessible authoring interface.

Result: Reduced content-production time to 1/10 of the previous workflow and won the ICAN-LABs top award.

Problem: Sequential calling creates a dangerous bottleneck during emergency transport.

Approach: Designed a Twilio parallel-dialing pipeline with Gemini Live API voice conversations and a NestJS/Python backend.

Result: Parallel dialing with automatic hang-up after acceptance, an estimated cost model of KRW 100–200 per case, and the 2026 Cheongryongthon Grand Prize (prototype, not deployed with real hospitals).

Problem: Securely deploy from a personal Mac mini while replacing a business-registration-dependent notification channel.

Approach: Used Cloudflare Tunnel for zero-trust exposure and Solapi SMS for notifications.

Result: Live service operation with PortOne payments integrated.

Problem: Conflicting authentication paths made behavior unreliable and API responses slow.

Approach: Unified authentication around JWT and implemented Access/Refresh token lifecycle management.

Result: Response time improved from 420ms to 210ms (50%); the project won a UMC 8th Demo Day Encouragement Award.

Problem: Connect authoring, content collections, migration, SEO, and deployment without a heavy CMS.

Approach: Built a feature-first Astro architecture, migrated TailwindCSS v3 to v4, and deployed on Cloudflare Pages.

Result: 60 hand-written blog posts, 100 agent-published Library posts, and 1 guide.

Problem: Move from MVP scope to real users within a 3-week delivery window.

Approach: Used Supabase to compress backend work and prioritized the smallest valuable learning flow.

Result: Deployed in 3 weeks and reached 100 weekly active users.