Blog

Notes from the forge.

What we read, what we built, and what we wish we'd known earlier. Researched essays on AI, workflow automation, and software for Indian businesses.

Index

Posts20
TopicsAI · Auto · Platform
CadenceWeekly-ish

AI Agents
20 May 2026 · 9 min
What agentic AI actually looks like in production
Most autonomous workflow demos collapse the moment money or compliance enters the loop. The realistic 2026 default is a hybrid, and the boundary line is the product.
- agents
- production
- governance
- hybrid
- operations
AI Agents
19 May 2026 · 9 min
MCP, explained for people who didn't read the spec
Anthropic's Model Context Protocol went from a niche RFC in late 2024 to the way every serious agent talks to its tools in 2026. Here's what it actually does, and where it still doesn't fit.
- mcp
- agents
- protocol
- tooling
- interop
AI Agents
18 May 2026 · 10 min
Why your agents need their own protocol to talk to each other
MCP handles what agents can touch. A2A handles how they coordinate. A field guide to LangGraph, CrewAI, AutoGen, and when multi-agent is actually worth the complexity.
- multi-agent
- a2a
- langgraph
- crewai
- orchestration
Workflow Automation
17 May 2026 · 9 min
Hyperautomation in 2026: what changed when AI ate the integration layer
RPA was screen-scraping. Hyperautomation is intent-routing across systems via AI. The new bottleneck isn't connectivity. It's trust.
- hyperautomation
- rpa
- integration
- ai
- smb
Workflow Automation
16 May 2026 · 10 min
n8n vs Make vs Zapier in 2026, with the honest tradeoffs
A working studio's view of all three. Zapier for time-to-value, Make for branching logic, n8n for sovereignty and 10x cost compression. The fit is the answer.
- n8n
- zapier
- make
- automation
- self-hosting
Workflow Automation
15 May 2026 · 9 min
Self-hosting n8n on a VPS: a pragmatic guide for SMBs
Why you'd self-host, what it costs (₹500 a month and an hour), the five things that go wrong, and when you should stay on cloud n8n instead.
- n8n
- self-hosting
- vps
- docker
- ops
Engineering
14 May 2026 · 10 min
LangGraph: when the complexity actually pays off
LangGraph is the most powerful and most painful agent framework. A walk through when state machines and checkpoints earn their cost, and when you should just use the Claude Agent SDK and move on.
- langgraph
- agents
- state-machines
- engineering
- framework
India SMB
13 May 2026 · 10 min
Workflow automation ROI for Indian SMBs: real numbers, not hype
82% of small business employers invested in AI in 2026. We break down what that money actually buys in India, with real payback numbers from our engagements.
- smb
- india
- roi
- automation
- gst
India SMB
12 May 2026 · 9 min
AI for GST compliance: from 12 hours to 2 hours a month
GSTN APIs, e-invoice IRP, and reconciliation finally became tractable for AI in 2026. What it cleans up, what still needs human eyes, and how to wire it without breaking your CA.
- gst
- compliance
- tally
- automation
- india
Engineering
11 May 2026 · 9 min
The cost-aware LLM pipeline: when to use Haiku, Sonnet, Opus
70% of LLM cost is wasted on calls that didn't need the smartest model. A working pattern for routing prompts to the right tier, with caching and graceful degradation.
- llm
- cost
- claude
- routing
- caching
Engineering
10 May 2026 · 10 min
RAG, fine-tuning, or just better prompts: a 2026 decision tree
Most teams reach for RAG when prompting would do, and fine-tuning when RAG would do. The honest decision tree, with the cases where each actually wins.
- rag
- fine-tuning
- prompts
- llm
- decisions
Engineering
09 May 2026 · 9 min
Vector databases in 2026: pgvector, Pinecone, Weaviate, Qdrant
Most teams should start with pgvector and not move until they hit a real wall. An honest comparison of when each one wins, with the operational tradeoffs.
- vectordb
- pgvector
- pinecone
- qdrant
- rag
Engineering
08 May 2026 · 9 min
Prompt engineering for production, not the chat window
Most prompt guides optimise for one-shot demos. Production prompts handle ambiguous input, partial failure, JSON edge cases, length limits, and tool-call recovery.
- prompts
- production
- engineering
- llm
- patterns
Workflow Automation
07 May 2026 · 9 min
AI governance when you're not Fortune 500
Big companies have governance committees. Small companies have one founder and a Slack channel. Here are practical guardrails that scale down to a ten-person team without theatre.
- governance
- smb
- ai
- compliance
- operations
India SMB
06 May 2026 · 9 min
The Indian SMB sales stack is WhatsApp plus AI now
WhatsApp is the entire sales surface for most Indian SMBs. A working stack of WAHA, Razorpay, Tally, and an AI agent replaces three roles at ₹16-40K a month.
- whatsapp
- smb
- ai
- india
- automation
Platform
05 May 2026 · 10 min
Building multi-tenant SaaS on Odoo 18, the Anti-Corruption Layer way
Why Krypto Forge Platform uses Odoo 18 as the ERP backbone but wraps everything in our own ACL. The trap of coupling product code to Odoo's data model, and how to stay free.
- odoo
- saas
- architecture
- acl
- multi-tenant
Industry
04 May 2026 · 10 min
AI for a 200-year-old industry: what we learned automating textile
The Paraslace story. What's hard about textile software (units, widths, dye-lots, job-work GST), and what AI in 2026 lets us do that we couldn't do 18 months ago.
- textile
- manufacturing
- smb
- india
- case-study
Industry
03 May 2026 · 9 min
AI for accountants: from bookkeeping to advisory
Bookkeepers aren't being replaced. The 60% of their work that was data entry is. What that does to pricing, firm structure, and the advisory work that's now the actual product.
- accounting
- ca
- ai
- india
- industry
Workflow Automation
02 May 2026 · 10 min
Self-healing workflows: error handling for autonomous pipelines
A workflow that runs at 2 AM and a workflow that breaks at 2 AM are different products. Patterns for retries, dead-letter queues, idempotency, and LLM-specific failure modes.
- reliability
- workflow
- error-handling
- sre
- automation
Engineering
01 May 2026 · 10 min
Vibe coding with discipline: why AI-assisted dev isn't yolo coding
The 'vibe coding' meme makes AI look like the end of engineering rigour. The teams shipping fastest with Claude Code and Cursor have more tests, more types, more discipline. Not less.
- ai-coding
- claude-code
- tdd
- engineering
- workflow

Notes from the forge.

All posts

What agentic AI actually looks like in production

MCP, explained for people who didn't read the spec

Why your agents need their own protocol to talk to each other

Hyperautomation in 2026: what changed when AI ate the integration layer

n8n vs Make vs Zapier in 2026, with the honest tradeoffs

Self-hosting n8n on a VPS: a pragmatic guide for SMBs

LangGraph: when the complexity actually pays off

Workflow automation ROI for Indian SMBs: real numbers, not hype

AI for GST compliance: from 12 hours to 2 hours a month

The cost-aware LLM pipeline: when to use Haiku, Sonnet, Opus

RAG, fine-tuning, or just better prompts: a 2026 decision tree

Vector databases in 2026: pgvector, Pinecone, Weaviate, Qdrant

Prompt engineering for production, not the chat window

AI governance when you're not Fortune 500

The Indian SMB sales stack is WhatsApp plus AI now

Building multi-tenant SaaS on Odoo 18, the Anti-Corruption Layer way

AI for a 200-year-old industry: what we learned automating textile

AI for accountants: from bookkeeping to advisory

Self-healing workflows: error handling for autonomous pipelines

Vibe coding with discipline: why AI-assisted dev isn't yolo coding