Decisions

Updated

How the systems were shaped.

Context, alternatives, trade-offs, and outcomes — including whether I’d make the same call today.

What
Production decision knowledge base
Who
Technical visitors evaluating judgment
Why
Trade-offs beat slogans
Next
Lessons · Changed my mind

Why Cloudflare D1?

ContextSmall team, edge deploy for Kathmandu firms, relational data with migrations — not a key-value dump.

AlternativesManaged Postgres · Firebase · pure KV/document store.

DecisionD1 + Workers for register/matters/API; documents in R2; hot cache in KV.

Trade-offsSingle-writer realities; disciplined schema; different scaling model than Postgres.

OutcomeOne person can ship and operate NepalIPMS without a DevOps hire.

Same today?Yes — for this team size and desk workload. Revisit if write-heavy multi-region demands change.

Why exact matching ahead of fuzzy matching?

I originally believed fuzzy matching would improve trademark search. It did, until it quietly introduced ambiguity in production.

Context130K+ trademarks; searches during live client calls; false conflict is worse than a missed suggestion.

AlternativesFuzzy-first similarity · hybrid rank · external search service.

DecisionNormalized exact/prefix FTS first; expand later only under control.

Trade-offsLess “smart” UI; more ingest normalization work.

OutcomeTrust restored; query latency suitable for desk calls.

Same today?Yes. Exact-first remains the default for register trust surfaces.

Why this OCR pipeline architecture?

ContextBulletins vary by issue; bilingual text; staff cannot trust unverified machine output in a matter file.

AlternativesSingle “AI upload” button · third-party black-box OCR SaaS only.

DecisionStore → OCR → classify → map → human review → merge; confidence on every field.

Trade-offsMore components and latency; safer legal records.

OutcomeProduction intake inside NepalIPMS with HITL intact.

Same today?Yes — stages can improve; the gate must stay.

Why domain model over generic CRM?

ContextIP firms already work company → file → matter → asset.

AlternativesHorizontal CRM · generic project tool · custom sheets.

DecisionMirror the desk; attach deadlines/documents/litigation to matters.

Trade-offsHarder horizontal pitch; higher daily fit.

OutcomeProduction desks on nepalipms.com including firm users.

Same today?Yes for specialized practice software.

Why human-in-the-loop for AI features?

ContextAuto-merge into legal records creates silent liability.

AlternativesFully automatic merge · AI chat without system of record.

DecisionSuggestions always land in review; no silent writes to matters.

Trade-offsSlower demos; stronger production trust.

OutcomeOCR and draft features that survive real desks.

Same today?Yes — non-negotiable for systems of record.

Why edge-native over always-on servers?

ContextNo dedicated DevOps; downtime in renewal season is a product failure.

AlternativesVPS + Postgres · classic containers · serverless elsewhere.

DecisionCloudflare Workers + bindings; same person designs and deploys.

Trade-offsCPU/SQLite constraints become design inputs.

OutcomeOperational simplicity; production LegalTech stack.

Same today?Yes for this operating model.

Why migrate in desk slices?

ContextBig-bang data moves freeze desks and hide failure modes.

AlternativesOne weekend cutover · freeze writes until perfect.

DecisionMigrate by workflow slices aligned to how firms actually work.

Trade-offsLonger calendar time; less elegant diagrams.

OutcomeRegister and matters entered production without freezing practice.

Same today?Yes — see lesson.

Why content architecture before marketing chrome (ScholarQuest)?

ContextStudy-abroad journeys fragment across destinations, tests, and inquiry.

AlternativesBrochure site · marketplace chrome first.

DecisionIA for destinations → prep → inquiry; conversion for advisers.

Trade-offsLess decorative launch; clearer counselling path.

OutcomeLive platform at scholarquest.com.np.

Same today?Yes for education consultancy products.

ADR template · Patterns · Knowledge index