_{Switch language / 언어 전환}

_{Decision case studies, told as judgment / 의사결정 케이스 스터디}

Product Planning × AI-Augmented Engineering
I turn ambiguous business needs into testable plans, then design and build the working systems with AI agents to prove whether those ideas survive real-world constraints.
My background is in operations, business planning, and data analysis. While working across those roles, I became increasingly drawn to the gap between ambiguous business needs and technical reality, and started building systems myself to close it.
Today, I focus on turning vague requirements into testable PoCs, structured experiments, and clear Go / Drop decisions grounded in real-world constraints.

"Memory is not a log. Memory is compacted meaning."

I am not a core AI researcher working on low-level optimization or deriving algorithms from first principles. My strength is execution, structured experimentation, and product-minded technical thinking.
When I encounter a business bottleneck, I combine existing models, APIs, and algorithms to test whether an idea is actually feasible under real constraints. I build to uncover structural limits, and I use evidence to decide whether to iterate, pivot, scale, or stop.

• Build to Validate: I use prototypes and workflows to test whether an idea can survive real business constraints.
• Production-Aware Thinking: I treat AI systems as services that must work with cost, stability, and operational friction in mind.
• Data-Driven Decisions: If a system does not prove its value through measurable results, I document the failure and move on.
• Learn from Limits: Failed experiments, trade-offs, and dead ends are often the most useful inputs for better system design.

• AI Proof-of-Concepts (PoCs)
• Automation Workflows
• LLM Application Prototyping
• Data Pipelines
• Technical Feasibility Validation
• Product-Oriented Experiment Design

• 🧠 Agent Memory System
  MCP-based long-term memory architecture.
  Architecture: Separated the State of Truth (MongoDB) from the semantic retrieval layer (ChromaDB) to improve consistency and memory compaction.

• 🎯 Logo Segmentation Experiment Workbench
  A comparison workbench for warp correction, anchor generation, Grounding DINO + SAM segmentation, and post-processing.
  Focus: Separated localization, segmentation, and quality experiments so failed assumptions remain inspectable instead of disappearing inside one pipeline.

• 🧪 Assessment Spec Harness (PoC)
  CI for hiring assessments — it doesn't evaluate the candidate, it evaluates the assessment design itself, detecting mismatches between a public spec.md and a private rubric.md before candidates ever see them.
  Architecture: A deterministic validation core over an immutable, hash-anchored source snapshot, exposed through an agent-consumable CLI contract. The current multi-run proof uses offline deterministic_extraction plus mock semantic verification; live LLM SDK runners remain deferred.

I value the lessons learned from failed experiments as much as successful deployments. Below are projects where I tested architectural ideas, examined feasibility, and made data-driven decisions.

• 🧪 Q-PSA (Project Killed)
  Tested discrete perturbation for quantized LLMs to estimate layer importance.
  Decision: Killed the project after experiments showed it was ~1300x slower than the baseline and failed pruning validation.

• 🗺️ Circle-WFC (Architectural Pivot)
  Attempted to replace A* pathfinding with a geometry-guided Wave Function Collapse (WFC).
  Insight: Found the structural limit of local consistency in global pathfinding, then reframed the concept as a candidate corridor generator whose output would still need A* or JPS for the final path.

• ⚡ HW-WFC v2.9 (Feasibility Validated)
  Constraint-driven AI compiler scheduling R&D.
  Result: Matched Exact DP's optimum, validating algorithmic feasibility, but concluded the research after identifying hardware-backed cost-model calibration as the real production bottleneck.

• 🧩 Harness IR (Feasibility Study — Mixed)
  Tested whether provider-neutral Role IR lowering beats hardcoded prompt templates for structured extraction across multiple LLM backends.
  Insight: Found no clean IR win on hard distractor sets — renewal/penalty false positives became a shared failure family across both paths — and reframed the real next-phase lever as self-verification and critic loops rather than the lowering step alone.

For longer write-ups on troubleshooting, architectural decisions, trade-offs, and project context:
👉 Selected Project Details (PROJECTS.md)

Email: [email protected]
LinkedIn: entangelk