AT Classic resume

Open to remote & on-site roles

Arsalan Tahir. I ship AI products end to end.

Full Stack AI Engineer. AI agents, RAG pipelines, evals and guardrails, plus the production SaaS around them. 4 years shipping. Ask below and watch the system answer.

Ask about Arsalan

grounded in his resume & projects · English / اردو

online

Hi 👋 I'm the assistant built into this portfolio. Ask me anything about Arsalan's work. While I answer, the system panel shows your question moving through the system and then opens the architecture of the project it matched.

live request trace · waiting for your question
·Edge route
·Rate limiter
·Retriever
·LLM
·Stream

$ system --map rag · How I build AI: a grounded, measured RAG pipeline

↳ click any node to see what I built there

years shipping production
4+
happy clients
20+
eval metrics scoring my RAG pipeline
5
products shipped
10+

// ai engineering

AI that holds up in production

Retrieval, evals, guardrails and voice: how I take LLM features from demo to dependable. Each one is backed by code I've shipped.

pgvector · full-text · RRF

Hybrid retrieval that actually recalls

I run cosine search over normalized embeddings (HNSW) and PostgreSQL full-text search side by side, merge the two with reciprocal rank fusion, then rerank the candidates with the model before anything reaches the prompt.

-- reciprocal rank fusion of vector + keyword hits
SELECT id, SUM(1.0 / (60 + rank)) AS score
FROM (SELECT id, rank FROM vec_hits
      UNION ALL SELECT id, rank FROM fts_hits) h
GROUP BY id ORDER BY score DESC LIMIT 8;

shipped in

Sourcebeam

result

Relevant passages for both semantic and exact-match questions

// selected work

Products I've shipped

Screenshots from the live products, plus the architecture behind each one. Toggle “System map” to see how it fits together.

SkooliQ AI answering a Roman Urdu request with a school snapshot
SkooliQ AI answering a Roman Urdu request with a school snapshot

01 · AI-powered school management SaaS

SkooliQ

Founder & sole engineer · 2026

One platform for admins, teachers, parents and students, with SkooliQ AI built in. Staff and parents ask in English, Urdu or Roman Urdu, and the assistant turns that into 20+ school workflows, from fee status to at-risk students. It also covers attendance, automated fee billing and recovery, PDF report cards, chat, push notifications and a mobile app. Every school is isolated as its own tenant.

  • SkooliQ AI: Gemini structured intent extraction drives 20+ role-scoped workflows in English, Urdu and Roman Urdu
  • Human in the loop: fee generation, reminders and announcements run only after admin confirmation, through expiring single-use actions
  • Insights and drafting: fee defaulter risk, at-risk students, report-card remarks and question papers with answer keys
  • WhatsApp channel for parents with signed webhooks, plus automatic Gemini model fallback
  • Designed, built and deployed the whole product as the only engineer
  • Automated monthly fee slips as PDFs, fee recovery tracking and reminders to unpaid families
  • Real-time chat and role-targeted announcements between staff and parents
  • Multi-tenant NestJS + PostgreSQL backend shared by the web dashboard and React Native app
Next.jsReactNestJSPostgreSQLGeminiSocket.IOCloudflare R2React NativeWhatsApp API
Sourcebeam AI workspace sign-in
Sourcebeam AI workspace sign-in

02 · Citation-first multimodal RAG workspace

Sourcebeam

Full stack + AI · 2026

Ask questions about private PDFs and images and get streamed answers with page-level evidence. It uses hybrid pgvector and full-text retrieval with rank fusion, Gemini reranking, an OCR fallback, durable ingestion jobs, and defences against prompt injection.

  • Hybrid retrieval: pgvector cosine (HNSW) + PostgreSQL full-text with rank fusion
  • Gemini reranking before grounded generation, with citations and match scores
  • OCR fallback for scanned PDFs; durable background ingestion with progress
  • Streaming NDJSON with cancellation, server-side sessions, rate limits, tests
ReactFastAPIPostgreSQLpgvectorGeminiSQLAlchemy 2.0TesseractS3
Code

↳ click any node to see what I built there

03 · Local speech-to-speech AI agent

Offline Voice Ordering Agent

AI engineer · Python · 2026

A coffee-shop kiosk that takes orders by voice with no cloud APIs, accounts or subscriptions. The customer speaks, the agent transcribes, reasons about the order with a local LLM, talks back to confirm, and logs the order for staff. After a one-time model download it works with no internet.

  • Speech-to-text with faster-whisper running on CPU
  • Conversation and order state handled by a small local LLM through Ollama
  • Neural text-to-speech with Piper for spoken confirmations
  • Fully offline after setup: private, free to run, no network latency
Pythonfaster-whisperOllamaPiper TTSLocal LLM
Code
RestaurantOS operations dashboard
RestaurantOS operations dashboard

04 · Restaurant operating system

RestaurantOS

Full stack: web, admin & API · 2026

One platform for POS, kitchen display, tables, delivery, riders, inventory, finance and a customer storefront, with an AI engine across all of it. It includes a kitchen voice agent that reads new orders aloud in English or Urdu, and a separate Super Admin portal for tenants, onboarding and subscriptions.

  • Seven role-specific apps: console, counter, kitchen, waiter, rider, customer, super admin
  • Kitchen Display System with a live voice agent and AI ETA predictions
  • NestJS modular monolith with Socket.IO, Prisma and PostgreSQL
  • Shared typed contracts package used by the API and every frontend
React 19ViteNestJSPrismaPostgreSQLSocket.IORechartsVoice AI
Extsy home: wallet, trading and P2P balances
Extsy P2P trading
Extsy trading screen
Extsy screen

05 · Crypto trading exchange

Extsy

Full Stack Engineer · Nextsense Solutions · 2025 · Shipped on iOS and web

A crypto exchange with spot, futures and demo trading, deposits, withdrawals, transfers, P2P, copy trading and admin tools. I built product features across the stack and worked on backend performance for 5,000+ users.

  • Cut API latency by 30% with query optimisation and leaner API workflows
  • Redis caching and rate limiting on the highest-traffic endpoints
  • Moved emails, notifications and AI calls to BullMQ queues with retries
  • Features across spot, futures, P2P, copy trading and admin tooling
ReactNext.jsNestJSPostgreSQLRedisBullMQTailwind CSSDocker

↳ click any node to see what I built there

06 · Computer vision · measurement from one photo

Object Size Measurer

AI engineer · Python · 2026

Estimates the real-world width and height of objects in a single photo, with no depth sensor. A reference object of known size calibrates the pixel-to-unit scale, then every other object's outline is measured and annotated.

  • Contour detection to find and order object outlines
  • Pixel-per-unit calibration from a reference object (for example a card)
  • Annotated output image plus printed dimensions
PythonOpenCVNumPyComputer vision
Code
BigTopa landing page
BigTopa landing page

07 · Virtual events & communities

BigTopa

Full Stack Developer · O'Cyber LLP · 2024

An interactive virtual-space platform for engagement, marketing and monetisation. I built the real-time communication and online-event features: video rooms, webinars and live broadcasts, all running on the LiveKit/WebRTC SDK I wrote.

  • Video rooms, webinars and live broadcasts on the in-house LiveKit SDK
  • Frontend components, backend APIs and LiveKit integration through to production
  • GraphQL APIs for real-time communication and data exchange
ReactNext.jsNestJSMongoDBLiveKitWebRTCGraphQL
Visit

// deploy log

Experience

Four years, three companies, every layer of the stack.

● running · Dec 2024 – Present

Full Stack Engineer @ Nextsense Solutions

  • Built an AI document-extraction pipeline that turns uploaded documents into validated structured data, using LLM structured outputs, schema validation and BullMQ background processing.
  • Built a Google Gemini chatbot with conversation memory, streaming responses and answers grounded in application data.
  • Built and optimised NestJS microservices for a crypto exchange with 5,000+ users, cutting API latency by 30% through query optimisation and leaner API workflows.
  • Added Redis caching and rate limiting to high-traffic endpoints, and moved slow work (emails, notifications, AI calls) into BullMQ background queues with retries.
  • Implemented RBAC, JWT/OAuth authentication and permission-aware APIs across multiple client products.
  • Delivered an asset-swap feature for XRP Healthcare end to end: swap and reward-distribution services, React web UI, React Native screens, and external wallet API integrations.
  • Implemented three-tier Stripe subscription billing for a screen-recording SaaS: checkout, webhooks, upgrades/downgrades, plan-based feature access.
  • Set up direct-to-storage media uploads with Cloudflare R2 presigned URLs, and Docker-based CI/CD for production apps.
#NestJS#Redis#BullMQ#Stripe#React Native#Docker

✓ deployed · Nov 2023 – Dec 2024

Full Stack Developer (MERN) @ O'Cyber LLP

  • Built a reusable LiveKit/WebRTC communication SDK for 1:1 calls, group meetings, webinars and live broadcasts, used across company products including BigTopa.
  • Developed Rugby.xyz player management, live scoring and performance dashboards with React, Node.js and PostgreSQL.
  • Improved API response times with query optimisation and indexing; implemented JWT auth, session management and RBAC.
  • Implemented GraphQL APIs and backend integrations for real-time communication and data exchange.
#LiveKit#WebRTC#GraphQL#PostgreSQL#Node.js

✓ deployed · Dec 2022 – Oct 2023

Frontend Developer @ Binary Marvels

  • Turned Figma designs into responsive, production-ready React applications, from build through deployment.
  • Implemented Google and LinkedIn OAuth and real-time collaboration features using WebSockets.
  • Built accommodation-management and event-participation platforms integrating third-party REST APIs.
#React#OAuth#WebSockets

// services: all healthy

Stack

AI engineering

up
RAGpgvectorEmbeddingsHybrid search + RRFRerankingRetrieval evalsPrompt-injection guardrailsTool callingWhisper / Piper / OllamaOpenCV

Languages

up
TypeScriptJavaScriptSQLPython

Backend

up
Node.jsNestJSExpressFastAPIRESTGraphQLMicroservicesSocket.IO

Data & performance

up
PostgreSQLMongoDBSupabaseRedisBullMQpgvectorIndexing

Frontend

up
ReactNext.jsReact NativeTanStack QueryRedux ToolkitTailwind CSS

Cloud & DevOps

up
DockerAWS EC2 / S3Cloudflare R2CI/CDLinux

Auth & payments

up
Multi-tenancyRBACJWTOAuth 2.0Stripe

AI & real-time

up
ClaudeOpenAIGeminiTool callingRAGLiveKitWebRTC

// engineering playbook

How I make it production-grade

The patterns I use again and again, with where they shipped and what they changed.

multi-tenancy · RBAC

Tenant isolation by default

Every request is resolved to a tenant and a role in a guard before it reaches a handler. Repositories only accept a tenant-scoped context, so an unscoped query can't be written by accident.

@UseGuards(JwtAuthGuard, TenantGuard, RolesGuard)
@Roles("admin", "teacher")
@Get("students")
list(@Tenant() t: TenantCtx, @Query() q: ListDto) {
  // every repository call is scoped to t.schoolId
  return this.students.findMany(t, q);
}

shipped in

SkooliQ · RestaurantOS · Nextsense client products

result

Every school and restaurant isolated on one shared deployment, with per-role access to every endpoint

$ sudo hire arsalan

Need someone who can own a feature from schema to production?

Open to remote & on-site roles. Based in Rawalpindi / Islamabad, Pakistan. Comfortable in remote, async teams.