Joshua Cho

Problem

LLM workflows were shipping into production for 10,000+ users, but latency, failure modes, and release risk were still treated like research experiments — not multi-team product systems.

Ownership

Owned backend services, public APIs, developer tooling, evaluation pipelines, and the CI/CD path that got AI features to production. Partnered with product, engineering, and research on capability design — not just prompt iteration.

Decisions

• Treated generation as a production pathway (API → model → eval → deploy) with explicit latency budgets, instead of optimizing prompts in isolation.
• Invested in monitoring and evaluation before feature velocity, so reliability gains stuck after each release.
• Automated cloud provisioning and deploy pipelines to remove release ceremony as the bottleneck on iteration speed.

Approach

Hardened scalable services and REST surfaces, reduced path length where it hurt p95, and made operability part of the platform — so research-driven features could move without weekend fire drills.

Outcomes

• Supported 10,000+ users in production
• API latency reduced 30%
• System availability up 20%
• Deployment time cut 40%
• Feature iteration speed up 25%

Stack in context

Backend services · REST APIs · Cloud infrastructure · CI/CD · LLM / RAG systems

Problem

Enterprise ordering needed higher engagement and lower ops cost, but slow release cycles and fragile backend/data paths made every improvement expensive.

Ownership

Led full-stack delivery as project lead: architecture decisions, API and backend design, cloud/DevOps automation, code review standards, and alignment with product and business stakeholders.

Decisions

• Chose service and data boundaries that reduced operational toil instead of bolting features onto a fragile monolith path.
• Prioritized deploy automation and workflow optimization together — faster releases without raising incident risk.
• Made architecture and review rituals explicit so quality scaled with pace, not against it.

Approach

Rebuilt the delivery loop around reliable APIs, database/workflow tuning, and cloud infrastructure that operators could actually run — then led the team through it.

Outcomes

• User engagement up 30%
• Operational costs down 20%
• Deployment cycles accelerated 50%

Stack in context

Full-stack · REST APIs · Cloud · DevOps · Databases

Problem

Mission-critical web systems needed better reliability and performance while modernizing legacy paths under strict security and quality constraints — change without fragility.

Ownership

Designed and delivered full-stack applications, backend services, REST APIs, and database work for secure enterprise environments. Drove legacy modernization and tighter development/deployment workflows without relaxing gates.

Decisions

• Modernized incrementally — replace high-risk legacy paths first rather than a big-bang rewrite under security pressure.
• Kept security and quality standards as invariants; optimized throughput by streamlining approved workflows, not by skipping them.
• Measured reliability and processing gains so modernization had operational proof, not just new code.

Approach

Ship secure applications and backends that operators can trust, then peel legacy debt in slices that survive audit-level scrutiny.

Outcomes

• Reliability and performance improved 25%
• Processing performance up 30%
• Delivery cycles reduced 20%

Stack in context

Full-stack · Secure APIs · Databases · Legacy modernization

Problem

Users needed dependable tracking on unstable networks. Lab latency numbers meant nothing if sync failed in the field.

Ownership

End-to-end product and technical ownership: mobile client, backend services, sync/fallback design, and performance iteration from live community feedback at 10,000+ daily users.

Decisions

• Designed for partial connectivity first — sync and fallbacks as core architecture, not an afterthought.
• Separated strong consistency needs from reconciliable state so the product degraded gracefully.
• Held a sub-300ms healthy-path target while optimizing for worst-case networks, not best-case demos.

Approach

Built resilient client/server sync with clear failure modes, then used feedback at scale to tune what users actually felt.

Outcomes

• Sub-300ms responses on stable networks
• Reached 10,000+ daily users

Stack in context

Mobile · Backend services · Resilient sync