RAG

How I Built a Multi-Tenant Legal RAG Platform That Saves Attorneys 30 Hours a Week

Multi-tenant RAG with hybrid search, tenant isolation, cross-encoder reranking, and citation grounding — and the 340-case golden-set eval that gates every deploy.

Multi-Agent

Orchestrating 50+ LLM Agents: Architecture, Cost Control, and Failure Modes

Hierarchical 50+ agent orchestration with per-agent context budgets, caching, and verifier-checked failure handling. What broke and what held.

Infrastructure

Why I Built a ₹10L GPU Cluster Instead of Paying for Cloud GPUs

Dual RTX 5090 local cluster beats cloud GPU costs within two years — the math, the tradeoffs, and what I'd do differently.

Career

From Commerce to AI Engineering: What 7 Years of Self-Teaching Taught Me

M.Com to Forward Deployed AI Engineer. Why shipping real projects beats tutorials — and what I'd tell my past self.

Retrieval

The Legal LLM Retrieval Problem: Why Hybrid Search Beats Semantic-Only

BM25 + dense embeddings + a reranker lifted legal retrieval accuracy from 71% to 89%. Why semantic-only is not enough.

Multi-Agent

Swarm OS: Building a Hierarchical Orchestrator-Worker Topology for Multi-Model AI

A 50+ agent, 5-model harness with context budgeting and cycle detection — and the failure modes that forced the design.

Automation

How I Use N8N + LangChain Together: 78 Workflow Templates Later

Where N8N stops and LangChain starts. 78 reusable templates across 10 domains, and the patterns that recur.

Deployment

Building AI Systems for Non-Technical Users: Lessons from 50+ Client Deployments

Why adoption — not accuracy — is the hard part. WhatsApp-first UX, voice, and meeting users where they already are.

Case Study

OpenSalesAI: Building a LangGraph Supervisor Agent for FMCG Distribution

A supervisor agent that detects intent and routes to 5 specialists over a 3-tier LLM fallback chain. Architecture, code, and the eval.

Evals

The Evals Gap: Why Most AI Engineers Are Flying Blind

Golden-set evals and LLM-as-judge gating — why they matter, how I build them, and the deploy-gate workflow that stops regressions.

Inference

Running vLLM + Ollama Together

Serving Qwen-72B INT4 with vLLM while keeping Ollama for the long tail of small models. Coexistence, routing, and latency.

draft Agents

Agent Harness: Turning Claude Code Into an Autonomous Development Operator

Wrapping Claude Code into a multi-agent harness — context budgeting, cycle detection, and staying inside the rails.

Fine-Tuning

Fine-Tuning Qwen/GLM with QLoRA: What Actually Moved the Needle

QLoRA on the 2× RTX 5090 rig — what I tuned, what I didn't, and when fine-tuning beats prompt engineering.

draft Deployment

Building AI for 8 Industries: What Generalizes and What Doesn't

Legal, finance, manufacturing, retail, agriculture, healthcare, e-commerce, education — the patterns that cross industries and the ones that don't.

Growth

MyKiranaBuddy: Scaling to 5,000 Users on ₹0 Marketing

B2B wholesale commerce, a field sales app, and a revenue engine I owned end to end — 5,000+ users, ₹50L revenue, no paid ads.

MCP

MCP in Production: Connecting Tools to Agents Without the Chaos

Model Context Protocol beyond the demo — production tool wiring, permissioning, and the failure modes that bite.

draft Career

The 30-Second Resume Test: How Recruiters Actually Read

What a recruiter sees in the first 30 seconds — and how to make those seconds count, from someone who's been on both sides.

Strategy

Why "Horizontal Solutions" Beat "Vertical Solutions"

Building one platform that adapts across 8 industries vs. eight point solutions — the economics and the engineering argument.