Insights
Writing
Technical writing and case studies by Arup Kanti Dash, focused on production systems, architecture decisions, and applied AI.
Decision-led case studies live on this site. Where a matching public artifact exists, each page links to the original LinkedIn argument, source code, and live product. Older essays still open on LinkedIn.
Architecture decisions
Decision records for the live Portfolio Ops system: why each choice, what I rejected, and the costs I took on. Each is verifiable in the running system. Read all →
- ADR-001 Pinecone serverless for vector search, not pgvector or OpenSearch
- ADR-002 A relevance floor over dense retrieval; the hybrid it called for shipped as ADR-011
- ADR-003 Amazon Nova Pro + Titan Embed v2 on Bedrock
- ADR-004 A bounded agentic loop with a self-check, not an open-ended ReAct agent
- ADR-005 Session-scoped, TTL-expiring, quota-bounded visitor uploads
- ADR-006 The corpus panel is generated from the live index, credibility is the product
- ADR-007 Defense in depth against prompt injection (OWASP LLM01)
- ADR-008 Deterministic fan-out for questions about Arup, not model-driven multi-hop
- ADR-009 Layered spend caps against denial-of-wallet, not a single limit
- ADR-010 Deterministic concept routing for abstract project questions
- ADR-011 Migrate from Pinecone dense-only retrieval to Qdrant hybrid retrieval
- ADR-012 Treat Qdrant hybrid retrieval as the current and future architecture baseline
- ADR-013 Disable the About Arup retrieval shortcut while keeping the profile fan-out
- ADR-014 Split ephemeral EKS into keep vs destroy stacks
- ADR-015 Chat mode: one durable transcript, never combined with Bedrock's own memory
- ADR-016 Push live quota over AWS IoT Core, not a held Vercel connection
- ADR-017 Breadth questions are enumerated, not searched
Case studies
- One Extra Correct Answer Is Not a Model Win
Close a real open-weight model gap on a free Colab T4, publish a reproducible adapter, and report what happened without turning one extra correct answer in a 30…
- Frameworks Are Not the Architecture
Learn LangChain, LangGraph, CRAG, and GraphRAG without creating a toy benchmark that makes every new technique look better than it is.…
- The Model Interprets; Code Computes
Turn a sentence, voice note, or booking screenshot into an India-specific comparison without allowing a fluent model to invent a carbon number that looks author…
- A 1:1 Tool Can Become Surveillance Software
Use session history to help an engineering manager prepare better 1:1s without quietly changing private notes into an employee-visible record or letting one rol…
- Neutrality Is a Build Constraint, Not a Prompt
Compare two fictional election manifestos in a way that teaches issue-level trade-offs without letting candidate identity, prompt order, or a fluent model decid…
- The Chatbot Was Harder Than the Queue Algorithm
Recommend an entry gate for a 50,000-person venue while queue load changes in real time, without creating herd movement or allowing a conversational model to in…
- The Backend Was Fast; the Product Still Felt Broken
An investment deal room received live updates while open, yet returning from the background could still show stale commitments for three to four seconds at the …
- Webhooks for Speed, Reconciliation for Truth
Orders from more than ten commerce platforms eventually reached the OMS, but four-plus minutes of polling lag had become an accepted operational problem on a re…
- Multi-Tenant Migration with Progressive Cutover
Migrate 150 B2B SaaS tenants from legacy infrastructure to a modernized multi-tenant platform. 10M rows across 30+ tables with complex foreign key dependencies.…
- When Semantic Search Misses Proper Nouns
The live portfolio RAG demo returned "no context" for Narrow Gate even though the phrase existed in an indexed upload. The old dense-only retrieval path ranked …
- When My Own Agent Said I Had No Weaknesses
A visitor asked the live demo "What are Arup's weaknesses?" and it answered that the corpus held no information about any, even though I had written real failur…
- Proving the Agentic Loop on Real Kubernetes, Not Just a Hosted API
The live Bedrock RAG demo proved a bounded agentic loop against a hosted model API. The open question was whether that same discipline holds up on infrastructur…
- When Agentic Retrieval Needed Determinism
The live portfolio demo failed a question it should have answered: "Is there an example of race condition handling, and what is its relation to idempotency?" Th…
LinkedIn essays
Essays that do not yet have a first-party case study. Where a website write-up exists, it appears once in the case-study list above.
- Choosing an AI Workflow Architecture: A Decision Example
AI Workflows · B2B SaaS
A decision framework for choosing RAG vs tools vs workflows under real constraints, complexity, cost, latency, and accuracy.
Related: AI Shopping Assistant · AI Itinerary Engine
- AI Won't Replace Your Learning. It Will Compress It.
AI Systems · Career Growth
How AI accelerates learning by compressing complex information, and why foundational principles like RAG and embeddings still matter.
Related: System Design Documentation · AI Architecture Strategy