Case Studies — Ajay Chaudhary

Legal services · India-based firm serving overseas clients

A growth engine built for the AI-search era

AI-search visitors convert ~10× better than classic organic — and this program is built around that

A law consultancy serving clients across the US, UK, Canada, and Australia needed enquiries, not reports. I run their entire web presence as a system: the site, a custom CMS, the analytics, and a growth program that ships improvements weekly behind a client-approval gate.

What was built

Technical SEO at scale: a 196-page site carrying 1,000+ structured-data blocks (FAQ, LegalService, professional credentials), 9-locale hreflang, a redirect consolidation architecture, and Core Web Vitals budgets enforced in CI.
AI-search (GEO) optimization ahead of the market: llms.txt, AI-crawler strategy, answer-capsule content structure, and citation tracking across ChatGPT, Perplexity, and Google AI Overviews — because that's where the highest-intent visitors now come from.
A self-built SEO intelligence platform: 20 data collectors (Search Console, GA4, Bing, rank tracking, competitor crawling), a 460,000-row data warehouse, and automated PDF + email reporting — running on a daily schedule, unattended.
Regulated-industry compliance: legal-advertising rules are enforced in four independent technical layers, including a server-side gate — no visible copy ships without written client approval.
Release engineering: every deploy is test-gated, and every changed URL is verified live after deploy, automatically.

Live in production — the program ships real changes weekly; client sign-off built into the pipeline
Measured, not claimed — every recommendation is backed by the platform's own warehouse, and outcomes are tracked against a weekly scorecard
Client reporting in plain English — automated summaries after every change batch, written for a non-technical owner

Stack: HTML/PHP on managed hosting · Python (20+ API integrations) · GitHub Actions CI/CD · GA4 · Search Console · DataForSEO

Education · residential school + exam-prep academy

Full platform: public site, student portal, and an AI content pipeline

~85 pages · ~800 automated tests · live on Cloudflare

A school and its exam-prep academy needed their entire web presence rebuilt: a public site with full content parity to the legacy CMS, plus a login-gated student portal with mock exams — and an AI pipeline to replace hours of manual content preparation every day.

What was built

Two Next.js applications deployed to Cloudflare Pages on every push, gated by typecheck and the full test suite.
Student portal: admin-approved signup, email-OTP two-factor for admins, timed mock exams graded server-side with negative marking, analytics, leaderboards, and a rank predictor.
Security designed into the database: row-level security on all 41 tables means students structurally cannot read answer keys or forge scores — enforced by triggers and proven by database-level regression tests, not by application code alone.
Browser-based exam proctoring using on-device face detection — no video ever uploaded, zero inference cost — detecting absence, multiple faces, and tab switches as advisory signals, never auto-punishment.
AI content pipeline: a newspaper article goes in; an exam-ready passage, five vetted multiple-choice questions, a printable worksheet, and vocabulary flashcards come out. Model output is schema-validated before it can touch the database, answer keys are permuted, and generation is refused unless a human approved the source text.
Prompt-injection resistance was tested against the live model — and when re-injecting rules proved insufficient, the defense was rebuilt structurally. Tested, not assumed.

Live in production — both apps deploy automatically; the academy uses the AI pipeline for daily content
~800 unit/component tests + database-level assertions — a pre-commit gate requires changed behavior to ship with tests
Cost-engineered AI — the full article-to-materials pipeline runs for a fraction of a cent per article, metered per job

Stack: Next.js 16 · React 19 · TypeScript · Supabase (Postgres, RLS, edge functions) · DeepSeek · MediaPipe · Cloudflare Pages

SaaS · own product

RAG chatbot SaaS: answer quality proven on 66 real businesses

66 help centers evaluated · thousands of scored Q&A pairs

A multi-tenant AI support chatbot I built and ran in production: a business points it at their website, documents, or Shopify catalog and gets an embeddable chat widget (works on any site) or a WhatsApp Business channel that answers from their own content — with citations.

What was built

Retrieval pipeline: hybrid vector + keyword search over the client's content, reranking, top-passage grounding, streaming answers with confidence scoring and source citations.
Quality engineering most chatbot builders skip: an evaluation harness that crawled 66 real company help centers (finance, e-commerce, legal, healthcare), generated questions, answered them through the live pipeline, and scored every answer with an LLM judge — producing per-section accuracy reports. That's how weaknesses were found and fixed before customers hit them.
14+ content source types: website crawl, sitemaps, PDF/DOCX, Google Drive, Notion, Zendesk, Shopify with live webhooks.
Production SaaS plumbing: multi-tenant auth with two-factor, subscription billing with metered usage, rate limiting, encrypted third-party tokens, error tracking, nightly offsite backups, and a disaster-recovery runbook.
Content-safety layer: crisis escalation, professional-advice disclaimers, and prompt-injection detection — deliberately rule-based so it can't be jailbroken.
Own ML where it counted: fine-tuned embeddings and a trained reranker from self-generated evaluation data.

Ran in production — multi-service deployment with custom domains, Postgres + Redis, monitoring and offsite backups
Evaluation as a discipline — the eval harness is reusable: it can score any AI system's answers against its own source content
Integration breadth — WhatsApp Business API, Shopify, Zendesk, Notion, Google Workspace, Stripe

Stack: TypeScript · Hono · Next.js · PostgreSQL + pgvector · Redis + BullMQ · DeepSeek/OpenAI · Stripe · Meta WhatsApp Cloud API · Railway

Three systems in depth

A growth engine built for the AI-search era

What was built

Full platform: public site, student portal, and an AI content pipeline

What was built

RAG chatbot SaaS: answer quality proven on 66 real businesses

What was built

Tell me what you're building