Abdullah Ghani

Backend Lead · AI Systems · Bengaluru

Abdullah Ghani

I build the backend — and the AI layer on top of it.

9+ years designing scalable APIs, job queues and cloud infrastructure. I lead backend at Delightree, where I built an AI Analyst agent that answers franchise owners' questions with verified numbers, and a RAG search engine over their SOP libraries.

Download resume See AI work

Years in backend engineering

Years architecting Delightree's platform

AI products live in production

Paid clients in 8 months at my startup

01 — AI Work

Production AI, built on solid backends

I treat LLMs as one more unreliable dependency: ground them in your data, compute the numbers in code, scope them by permissions, and keep them off the request path. Here's what that looks like in practice.

Delightree · LLM agents · LangChain · LangGraph

AI Analyst — an agent franchise owners can trust

Franchise owners ask questions in plain English, like "Which locations are falling behind on training, and why?", and get verified answers in seconds, with tables, charts and one-click actions. The hard part isn't the chat. It's making sure every number the agent shows is correct.

RouteAn LLM router selects the right tools from a catalog spanning the whole business

→

GatherA ReAct agent calls those tools against live platform data

→

ComputeEvery figure is calculated in code from the shared metric catalog, never guessed by the model

→

AnswerA second model writes a schema-validated answer that is checked before it is shown

Two-stage design: the router narrows a large tool catalog down to the few tools a question needs, which keeps the agent fast and focused.
Built on the LangChain ecosystem: the router → ReAct → answer flow is orchestrated as a LangGraph graph, with every data source exposed as a typed LangChain tool.
Numbers you can trust: results are checked again and rebuilt from the same metric catalog the dashboards use, so chat and dashboards always agree.
Structured output: answers follow a schema, so the frontend can show tables, charts and action buttons instead of plain text.
From answer to action: answers come with one-click actions, so owners can act on what they find without leaving the chat.

Delightree · RAG · Permission-aware retrieval

AI Search — cited answers from a company's own knowledge

Franchise teams used to dig through binders, PDF libraries and group chats to answer everyday questions like "What's the refund process for a delivery order?" or "What's the closing temperature for the line?" AI Search is a vector-based RAG engine that answers those questions in plain English, on any device, and shows the source SOP behind every answer so teams can check the original.

IndexSOPs, manuals, training modules, policies, checklists and audits are embedded into a vector index that updates automatically when content changes

→

ScopeRetrieval respects the network's permissions: franchisees only see their brand's content, and multi-brand operators control visibility per brand

→

AnswerThe LLM answers only from the retrieved, approved content and returns the exact answer with its source reference

→

LearnThumbs-up/down feedback and search analytics reveal knowledge gaps and trending topics for HQ

Grounded by design: answers come only from content the franchisor has uploaded and approved, so every location gets the same current answer.
Multi-tenant safe: retrieval is filtered by the same permission hierarchy as the rest of the platform, so brands never see each other's content.
No separate pipeline: it works on content already in Delightree in any format, and edits to the knowledge base reach search automatically.
Less load on HQ: locations get answers on their own around the clock, and new hires ramp up without waiting on a manager.

RAGEmbeddingsVector searchMulti-tenant access controlNode.jsGraphQL

Product page ↗

RAG · LangChain · Vector search · Background AI workers

Brifly — give your meetings a memory

Meetings get recorded in Read.ai, Fireflies, Zoom and Microsoft Teams, and follow-ups happen over email, so nothing can be searched in one place. Brifly pulls all of it into one private library and adds an AI layer: ask questions across every meeting, chat with a single meeting, get weekly digests and turn action items into tasks. Every answer cites the exact speaker and timestamp.

IngestRead.ai (OAuth 2.1), Fireflies (GraphQL), Zoom & Teams (OAuth 2.0, VTT parsing), plus read-only Gmail & Outlook

→

QueueBullMQ + Redis workers: embed, enhance-actions, digest

→

IndexOpenAI text-embedding-3-small → Qdrant, with per-meeting and global collections

→

AnswerClaude Haiku 4.5, grounded in retrieved transcript chunks and citing speaker + timestamp

LangChain RAG pipeline: transcripts and emails are chunked, embedded and retrieved through LangChain, with Qdrant as the vector store and Claude generating the cited answers.
Two RAG modes: scoped retrieval for chatting with one meeting, and global retrieval across all meetings with per-source filters. Conversation history is saved and can be copied, regenerated or downloaded.
One schema for every source: four meeting providers with different auth flows and formats are normalised into one model. Split "Conference Room – Speaker 1/2" tracks are merged into one speaker timeline.
Action items to tasks: Claude adds priority, assignee and deadline to existing action items, or extracts them from raw transcripts. Tasks are grouped as overdue, today, this week or later.
Scheduled AI digests: weekly digests (Mondays at 8 AM in the user's timezone) and monthly ones cover key themes, decisions and the top action items, each linked back to its source meeting.
Email intelligence: read-only Gmail and Outlook sync finds the key threads, action items and who's waiting on your reply.
Built for production: sync runs sequentially with retry-with-backoff on 429s and concurrency-safe jobs, and LLM and embedding work stays off the request path. Auth uses email/password or Google OAuth, with accounts merged by email.

Node.jsExpressMongoDBBullMQRedisQdrantAnthropic ClaudeOpenAI EmbeddingsReactDocker

Live app ↗

Generative AI · Hosting infra

DriftHost — describe a website, watch it go live

An AI website builder and static hosting in one product. You chat, and the AI streams real HTML, CSS and JS while the live preview updates. You can click any element in the preview and say what to change, or edit the files directly. Every AI turn is checkpointed and can be reverted. Publishing creates an immutable, versioned snapshot on yourname.drifthost.site, served from an edge CDN, and rolling back takes one click. Developers can skip the chat and run drifthost deploy.

LLM codegenStreamingCheckpointsVersioned deploysEdge CDNCLI

drifthost.site ↗

02 — Experience

Where I've shipped

Backend leadership and senior engineering across franchise-tech, fintech, customer-experience analytics and identity verification. Bachelor's in Computer Science, Netaji Subhash Engineering College (2017).

Aug 2025 — Present

Backend Lead · Delightree Current

Franchise management platform. I lead backend development, with a focus on AI-powered features.

Built the AI Analyst, a two-stage LLM agent on LangChain and LangGraph. A router selects the right tools from a catalog spanning the whole business, a ReAct agent gathers live data, and a second model writes a schema-validated answer.
Made its numbers trustworthy: every figure is computed in code, then checked again and rebuilt from the same metric catalog the dashboards use, so chat and dashboards always agree.
Built AI Search, a vector-based RAG engine that answers plain-English questions across franchisors' whole SOP library and cites its sources.
Introduced an AI-SDLC process across the engineering org, building AI assistance into how we plan, build, review, test and ship software.

Aug 2021 — Aug 2025

Senior Backend Engineer · Delightree

Designed the backend architecture of the franchise platform on Apollo GraphQL, giving franchisors and franchisees one scalable API layer.
Built the core modules franchise teams use every day across locations: SOPs, Forms, Audits, Training and Notifications.
Built background job processing and caching with Bull queues and Redis, moving heavy work off the request path.
Deployed and ran the AWS infrastructure: Elastic Beanstalk, ECS, CloudFormation, S3 and CloudFront.

Sep 2022 — 2023

Senior Full-Stack Engineer · Chain of Demand, Hong Kong

Crypto investment platform with machine-learning-driven quantitative analytics.

Designed and built the backend APIs behind the core investment services.
Launched a signals platform that surfaces the best-performing ML strategies and alerts users in real time via Telegram.

Jun 2019 — Jul 2021

Co-Founder · DecodeCX

Customer-experience and online-reputation platform for local businesses.

Designed the product and built the platform from scratch, from MVP to production.
Built review-aggregation pipelines and aspect-level NLP that turn unstructured customer feedback into insights.
Reached 60+ paid clients within 8 months.

Jan 2018 — Jun 2019

Full-Stack Developer · Survaider

Customer-experience and feedback-analytics platform.

Built and maintained 20 web scrapers feeding the analytics pipeline.
Built an internal Operations Platform that saved the team ~150 work hours a month.
Built the end-user analytics platform used by ~800 B2B users.

Jun 2017 — Jan 2018

Software Developer · Signzy

Digital onboarding and identity verification for banks.

Built onboarding applications for major Indian banks on in-house KYC APIs.
Did research and development on OCR for the company.
Maintained in-house backend APIs on Node.js.

Aug 2016 — Jun 2017

Web Developer · ApnaStudy Technologies, Kolkata

Built the payments platform for the existing user base.
Fixed bugs and maintained APIs built on the MEAN stack.

03 — Projects

Other products I've built

Full-stack platforms in production: construction ops, mobility and services.

Crewd

All-in-one construction management for residential builders. Leads, jobs, schedules, budgets, change orders, invoices and daily logs live in one system, so data entered once flows to every later step.

GraphQLTypeScriptReal-time

crewd.site ↗