AI Backend Engineer

AI Backend Engineer

Backend systems for LLMs, RAG, and AI automation — in production.

Why hire me as AI Backend Engineer?

I build the backend that makes AI features actually ship: retrieval pipelines, semantic search, LLM integrations, and the async infrastructure around them. I've built crypto/social intelligence and AI content platforms with Google Gemini and OpenAI, semantic search over pgvector, and deployed local LLM inference (vLLM + LiteLLM) behind an OpenAI-compatible API for a team. I treat AI as an engineering problem — evaluation, fallbacks, caching, and cost — not a demo.

What I've shipped

  • Designed a crypto/social intelligence backend ingesting Telegram, Discord, and X, processed through AI summarization, embeddings, semantic search, and alerting
  • Integrated Google Gemini embeddings with pgvector semantic search and Redis caching/queueing for retrieval and background processing
  • Deployed local LLM inference with vLLM on constrained GPU hardware and configured LiteLLM for OpenAI-compatible team access
  • Built AI content pipelines with versioned prompt management (n8n) so prompt changes deploy without code changes
  • Built and open-sourced HireTrack — a deterministic-first LangGraph pipeline for résumé tailoring across 5 AI providers

Core stack for this role

Python FastAPI Django OpenAI API Google Gemini LangChain LangGraph pgvector Redis vLLM / LiteLLM Celery Docker

Relevant Projects

Work that directly demonstrates this specialisation.

HireTrack — AI Résumé Builder & Job Pipeline Tracker

Open-source, bring-your-own-key web app that tailors ATS-ready résumés to each job and tracks an interview pipeline — the tool I use to run my own job search. A deterministic-first LangGraph pipeline extracts job-description data and scores match/positioning; résumés export to ATS-safe DOCX, PDF, or Markdown entirely client-side. Truth-only tailoring mirrors JD phrasing with no invented skills or metrics, and every run appends honesty/verification notes. Supports 5 AI providers (Anthropic, OpenAI, Gemini, and any OpenAI-compatible endpoint), Supabase Postgres + Auth with row-level security, and installs as an offline-first PWA.

  • 5 AI providers (BYOK)
  • LangGraph JD pipeline
  • ATS-safe DOCX/PDF export
React 19TypeScriptViteLangGraphSupabasePostgreSQL

AI-Driven Cryptocurrency Intelligence Platform

Architected a real-time intelligence backend as sole engineer, ingesting social data from Telegram, Discord, and X and processing it with Google Gemini embeddings and pgvector semantic search. Built async multi-chain blockchain analysis services with token-bucket rate limiting and Redis caching, plus ML-based DEX scam-token detection. Orchestrated ~10 services via Docker Compose with migrations, background workers, and structured logging.

  • ~10 microservices
  • 6 chains analyzed
  • ~92% scam detection (test set)
PythonFastAPIDockerGoogle GeminipgvectorRedis

AI-Powered LinkedIn Content Platform

Primary DRF backend developer for a production AI-powered LinkedIn content platform. Built 7 backend services and 60+ REST endpoints, including versioned prompt management across AI content pipelines (n8n), multi-tier profile extraction (LinkedIn API → scraper → OpenAI fallback), real-time notifications via Django Channels WebSockets, Stripe-based subscription billing, OAuth2 + JWT refresh auth, and semantic search using Gemini embeddings + pgvector with Redis caching.

  • 7 backend services
  • 60+ REST endpoints
  • Gemini + pgvector search
PythonDjangoDRFWebSocketsOpenAIpgvector

Local LLM Inference Infrastructure

Deployed internal LLM inference for team use: vLLM on an RTX 3080 Ti (12 GB VRAM), with a LiteLLM proxy exposing a unified OpenAI-compatible API. Benchmarked local models under hardware constraints and selected Qwen3-8B-AWQ for coding and automation workflows, tuning context-window and prefix-caching settings. Wired OpenCode for agentic coding.

  • vLLM on 12GB VRAM
  • LiteLLM OpenAI-compatible proxy
  • Local model benchmarking
vLLMLiteLLMQwen3-8B-AWQPythonOpenAI-compatible APIGPU Inference

Cardiovascular Age Prediction — Published Research

End-to-end ML pipeline for cardiovascular age prediction from single-lead ECG. Collected data from 42 subjects using ESP32 + AD8232, producing 6,131 segments with 13 features. Trained Random Forest (R²=0.99) and ResNet-18 transfer-learning models (R²=0.87). Co-authored (equal contribution) and published in Circuits, Systems, and Signal Processing (CSSP, Springer), 2025 — DOI 10.1007/s00034-025-03048-2.

  • 42 subjects · 6,131 segments
  • R²=0.99 (Random Forest)
  • Published · CSSP (Springer) 2025
PythonKerasscikit-learnPyTorchESP32IoT

Need a tailored résumé?

This résumé is focused specifically on AI Backend Engineer work — highlighting the most relevant projects and skills.

Download AI Backend Engineer Résumé (PDF)

Frequently asked

What AI/LLM backend experience does Mubashir Rehman have?

Mubashir has built production AI backends including a crypto/social intelligence platform using Google Gemini embeddings and pgvector semantic search, AI content pipelines with versioned prompt management, and local LLM inference infrastructure with vLLM and LiteLLM. He also built HireTrack, an open-source LangGraph-based résumé-tailoring tool supporting 5 AI providers.

What is his core AI backend stack?

Python with FastAPI and Django, OpenAI API and Google Gemini, LangChain/LangGraph, pgvector for embeddings and semantic search, Redis for caching and queueing, Celery for background work, and Docker — plus vLLM/LiteLLM for self-hosted inference.

Have a specific question?

Ask directly →