Blog

Learned in Production. Written Down.

Practical articles on production AI, AWS architecture, and lessons from shipping.

Case Studies

Production systems we built. Each with deep dives into specific problems.

Case Study

An Airline in the Chat Window: Building an Agentic Commerce Accelerator on ACP and MCP

2026-08-2414 min read|1 deep dives · 27 min total

We built an agentic commerce accelerator that puts merchants inside ChatGPT and Claude: full ACP checkout sessions, delegated payments, and a 12-tool MCP server, with airline booking as the deepest vertical. Protocol-complete and demo-tested, not yet live: the architecture, the spec-versus-reality gaps, and what fronting a real carrier would take.

1.MCP Protocol in Practice

Read case study →

Case Study

Anatomy of a Production RAG System: AI-Powered Title Validation for Indian Real Estate

2026-06-1112 min read|5 deep dives · 59 min total

We built a document intelligence system that validates property title cleanliness across over a thousand properties in Indian metros, processing sale deed chains, encumbrance certificates, and revenue records in six languages. The architecture, what broke, and what we shipped.

1.Chunking Strategy for Production RAG 2.When Semantic Search Fails 3.Debugging Production AI 4.Guardrails for Production AI 5.The AI Security Questions Enterprise Buyers Are Starting to Ask, and How to Answer Them

Read case study →

Case Study

LangChain Agent to Production: WhatsApp Hotel Booking

2026-04-0310 min read|6 deep dives · 63 min total

A production conversational AI handling hotel bookings over WhatsApp: room availability, pricing, confirmation, and post-booking queries. Where cost surprises, retrieval failures at scale, and PII handling problems first surfaced.

1.LangChain and LangGraph in Production 2.Agentic Architecture Patterns 3.How SaaS Companies Should Be Billing for AI Features 4.Per-Customer AI Cost Attribution 5.How Infrastructure Decisions Determined Our AI Margins 6.Testing Conversational AI in Production

Read case study →

All Articles

Deep Dive

MCP Protocol in Practice: What We Learned Implementing Model Context Protocol for AI Commerce

2026-09-0313 min

MCP lets AI assistants interact with external systems. We implemented it for a commerce platform. Here's what the specification doesn't tell you.

MCPModel Context ProtocolAgentic AI

Case Study

An Airline in the Chat Window: Building an Agentic Commerce Accelerator on ACP and MCP

2026-08-2414 min

MCPACPAgentic Commerce

The Gap Between AI Demo and Production: A Checklist for What Changes

2026-08-209 min

The AI demo works. The board is excited. Now ship it to real users. Here is everything that changes between a proof-of-concept and a production AI system, from a team that learned most of these the hard way.

Production AIPoCMVP

Deep Dive

The AI Security Questions Enterprise Buyers Are Starting to Ask, and How to Answer Them

2026-08-1710 min

Enterprise deals are stalling on AI security questions most mid-market companies can't answer. Prompt injection, data leakage, agent vulnerabilities. Here's what we learned processing tens of thousands of documents containing Aadhaar numbers, PAN details, and bank account information.

AI SecurityPrompt InjectionOWASP LLM Top 10

Deep Dive

Testing Conversational AI in Production: DeepEval, Prompt Regression, and the Evaluation Stack

2026-08-1312 min

Manual prompt testing catches obvious failures and misses everything subtle. The evaluation stack that gave us confidence to ship model swaps, prompt updates, and retrieval changes to a live booking system.

DeepEvalTestingLLM Evaluation

Deep Dive

Guardrails for Production AI: Citation Checking, Confidence Thresholds, and Human Escalation

2026-08-1010 min

Guardrails are not prompt instructions. They are infrastructure. Citation checking to catch hallucinations, confidence thresholds to trigger human review, and escalation paths that keep AI reliable when the stakes are high.

GuardrailsReliabilityHallucination Detection

Deep Dive

How Infrastructure Decisions Determined Our AI Margins

2026-08-067 min

We did not plan infrastructure around AI economics. We built the agent, then discovered what was missing when margins we could not explain started compounding. The gateway, the PII vault, the scaling policy: each was forced by a failure, not a checklist.

AWSInfrastructureAI Margins

Deep Dive

When Semantic Search Fails: Building Hybrid Retrieval for Production RAG

2026-08-037 min

Pure vector similarity search works on demo datasets. In production, it returns plausible-sounding wrong answers for the queries that matter most: exact identifiers, proper nouns, reference numbers. Hybrid search with re-ranking recovers precision without sacrificing recall.

RAGHybrid SearchRe-ranking

Deep Dive

Debugging Production AI: The Observability Stack That Tells You Why Your System Broke

2026-07-279 min

When a production AI system returns wrong answers, traditional APM tells you the request succeeded. It did. The answer was just wrong. Here is the observability stack we built to catch quality failures before users reported them.

ObservabilityDebuggingLangfuse

Deep Dive

Per-Customer AI Cost Attribution: Building the Margin Map That Changed Our Pricing

2026-07-209 min

We had hundreds of hotel properties on two commercial models and no idea which ones were profitable. The per-tenant margin map - built on Langfuse, a LiteLLM-style gateway, and Stigg - was the most important tool we shipped.

AI CostCost AttributionPer-Tenant Economics

Deep Dive

Chunking Strategy for Production RAG: What We Changed After Real Documents Broke Ours

2026-07-1311 min

The most common root cause of production RAG failures is not the model. It is the chunking strategy. Fixed-size chunking that worked in demos breaks on legal documents, financial reports, and technical manuals. We learned this processing tens of thousands of Indian property documents.

RAGChunkingDocument Processing

The AWS Infrastructure Checklist We Run Before Shipping AI

2026-07-0913 min

Compute, storage, networking, security, observability, cost - the AWS checklist we run before an AI system ships. Some items we learned the hard way; the rest are defaults we now reach for before the failure arrives.

AWSInfrastructureChecklist

Deep Dive

Agentic Architecture Patterns: When to Use Multi-Step Agents, Tool Chains, and Orchestration Layers

2026-07-078 min

Not every AI problem needs an agent. But when it does, the architecture patterns matter. Four patterns from production systems, one anti-pattern we lived through, and a decision framework.

AgentsAgentic AIArchitecture Patterns

Deep Dive

LangChain and LangGraph in Production: What Works, What Breaks, and What We'd Change

2026-06-308 min

We shipped two production systems on LangChain and LangGraph. An honest assessment: what the framework handled well, where we fought it, and the patterns that survived.

LangChainLangGraphAgents

Deep Dive

How SaaS Companies Should Be Billing for AI Features: Metering, Entitlements, and the Tools That Exist

2026-06-229 min

Your AI features are live but billing hasn't caught up. The metering tools that exist, how to evaluate them, and patterns that work, grounded in real per-tenant economics from a production AI system.

AI BillingMeteringEntitlements

Case Study

Anatomy of a Production RAG System: AI-Powered Title Validation for Indian Real Estate

2026-06-1112 min

RAGDocument AnalysisReal Estate

Case Study

LangChain Agent to Production: WhatsApp Hotel Booking

2026-04-0310 min

LangChainLangGraphWhatsApp

Have a Problem Like the Ones We Wrote About?

Tell us where you are with AI. Straight feedback on what makes sense — and what doesn’t.

Let's Talk →