Architecture · Design Philosophy
Engineering Practice & Principles
Pragmatic, shippable, learned by building
A set of pragmatic engineering principles — from technology selection to testing to retrospectives. Not textbook theory: practiced and refined in recent hands-on projects like job-hunt and LearnArken.
Design Philosophy
Stability First
Choose production-ready, stable solutions (e.g., managed cloud services). Small teams cannot afford to spend time on solution stability issues.
Stand on Giants' Shoulders
Reference industry best practices. You're not the first to face this problem — leverage others' experience to go further.
Cost-Effectiveness First
Expensive solutions are the enemy of ROI. Use pay-per-use serverless in MVP phase, switch to fixed services when traffic is predictable.
Documentation First, Standards First
All scenarios should be considered before coding begins, not researched on the fly.
Testing Has Veto Power
No product ships without tests. Every change, even minor dependency updates, requires testing.
New Tech Must Be Targeted
New technology must solve a specific problem. Full replacement is nearly impossible. Research and deployment must have clear goals.
Redundant Monitoring
Never rely on a single monitoring channel — auto-call + auto-message. Production incidents happen more than expected; ensure minimal impact.
High Availability Is a Feature
HA cannot stay theoretical. Make it understandable to non-technical stakeholders and visibly part of the product.
Proactive Communication When Issues Arise
Unexpected things will happen. Communicate proactively — don't hide or sugarcoat. Conduct proper post-mortems.
Post-Mortem Is the Final Step
Development without reflection isn't development. A project without retrospective isn't a project.
Project Architecture Cases
Not just theory — real projects proving the design philosophy
job-hunt
Three subsystems · human in the loop
- ✓Evaluation and application decoupled, each testable
- ✓460+ pytest tests; testing has veto power
- ✓Fill-only; never auto-submits behind unmet gates
LearnArken
Advanced RAGHybrid retrieval → reranking → fail-closed grounding
Phase 1: Hybrid Retrieval
BM25 structure-aware chunking + dense vectors (Vespa) fused with RRF, then refined by a cross-encoder reranker
Phase 2: Grounding Gates
Three fail-closed gates and verbatim-quote validation — refuse rather than fabricate; citations backfilled mechanically from metadata
Phase 3: Adversarial Eval
A 32-case adversarial set scored by two heterogeneous judges; a cross-document aggregation hallucination found and fixed
RAG for safety-critical aviation docs: reproducible and auditable — only reproducible numbers enter the README