AI & Backend Engineer

I build LLM systems, and write about how they work under the hood.

RAG pipelines, agents, fine-tuned models and the backends that keep them fast. I'm working through transformers from first principles, in public.

Read the writingSee projects
Remote

Worked with teams in Canada, Singapore, the US and Germany

Investor search · Metal
500ms<50ms

pgvector search over 100K+ profiles

Self-attention, live
wewalkedalongtheriverbankwewalkedalongtheriverbank

bank looks most at river (47%) after itself. Hover a word to try another.

Latest article · November 2, 2026

Mixture of Experts for LLM Inference

Why 671B can run like ~37B per token: FFN as memory, routers and top-k, fine-grained experts, shared experts, routing collapse, and the all-to-all serving tax.

Read it →
Series · 7 parts

Transformers from first principles

From tokens and dot products to a trained GPT-2, one idea at a time.

1234567
Project · Felixai

Event-driven Canvas LMS sync

Event-driven platform syncing and processing Canvas LMS content. Apache Kafka pipelines for large-scale ingestion, OCR, and AI workflows; Server-Sent Events for real-time sync progress; tuned for I/O-bound performance.

Canvas→Kafka→OCR→LLM→SSE
Project · VecLite

A vector store on a K-means tree over SQLite

Efficient vector store using a K-means tree over SQLite. Cluster-based search, full scans, and Random Projection dimensionality reduction for fast retrieval.

PythonSQLite
Project · AutoGrad Engine

Automatic differentiation from scratch

A from-scratch automatic differentiation engine in C++ for efficient gradient computation.

C++
Pinnacle Bot
700pages

Civil engineering PDFs turned into structured documents

Experience
Full history ↓
  • Effica AI Engineer2025–26
  • Metal AI Backend Engineer2025
  • Mosaik AI Engineer2024–25
  • Beam ML / Backend Engineer2023–24
Stack
  • Python
  • JavaScript
  • TypeScript
  • Go
  • Node.js
  • FastAPI
  • LLMs
  • LangChain
  • LangGraph
  • Llama-Index
  • RAG
  • HuggingFace
  • Fine-Tuning
  • PyTorch
  • TensorFlow
  • Deep Learning
  • Object Detection
  • OpenCV
  • Vector Databases
  • PostgreSQL
  • pgvector
  • Docker
  • AWS (S3 / EC2 / Lambda)
  • REST APIs

01 · Systems

How I build

Dimensions in milliseconds

DWG-01 · FELIXAI INGESTIONEVENT-DRIVEN
Canvas LMSsourceKafkatopicsOCR workersI/O boundAI workflowsLLM callsPostgrescontent storeSSE streamprogress eventsBrowserclientASYNC · SCALES BY PARTITION12live sync %
Canvas LMS content flows through Kafka to OCR and AI workers, and sync progress streams back to the browser over Server-Sent Events.
DWG-02 · INVESTOR SEARCHMETAL
Queryembed()pgvectorsimilaritySQL filters15+ predicatesRankedtop-k investorsBEFORE · 500 msAFTER · <50 ms · 10× FASTER
Vector search in pgvector plus 15+ SQL filters, tuned from 500ms to under 50ms on 100K+ profiles.

All projects · 8

Felixai

Event-driven platform syncing and processing Canvas LMS content. Apache Kafka pipelines for large-scale ingestion, OCR, and AI workflows; Server-Sent Events for real-time sync progress; tuned for I/O-bound performance.

Kafka · OCR · SSE · Python

AI Calling Agent

Outbound voice-calling agent built with Vapi. Real-time STT/TTS pipelines for natural dialogue, with dynamic prompt logic and external tool integration for scenario-based workflows.

Vapi · STT/TTS

HailoChat

Multi-tenant Zendesk-style support SaaS. Scalable WebSocket chat with Redis-backed sessions, hybrid PostgreSQL + MongoDB storage, and subscription-based access control.

WebSockets · Redis · PostgreSQL · MongoDB

VecLite

Efficient vector store using a K-means tree over SQLite. Cluster-based search, full scans, and Random Projection dimensionality reduction for fast retrieval.

Python · SQLite

Email Intelligence System

FastAPI + Azure OpenAI service that extracts structured data (order/PO/registration numbers) from .eml files and attachments, cross-verified against a PostgreSQL inventory DB.

FastAPI · Azure OpenAI · PostgreSQL

Majai Platform

AI ad-automation for Facebook/Google Ads. Content generation backend with image/video models, Facebook Graph API auto-posting, ESRGAN enhancement, and an OpenCV video pipeline.

Graph API · ESRGAN · OpenCV

Pinnacle Bot

RAG-based document tool processing 100–700 page PDFs, generating structured documents for civil engineers with extracted insights and reference links.

RAG · PDF parsing

AutoGrad Engine

A from-scratch automatic differentiation engine in C++ for efficient gradient computation.

C++

02 · History

Experience

Newest first

  1. May 2025 — June 2026

    Effica

    · AI Engineer (Contract)

    Remote · Canada

    • Built a serverless document ingestion pipeline on AWS Lambda for large health documents and research articles, using a two-stage parser for structured extraction and pgvector for semantic search.
    • Developed an end-to-end health report generation system from doctor–patient calls: audio → transcription → LLM/NLP pipelines producing multiple clinical and practitioner-ready report types.
    • Built prompt-configurable report workflows letting practitioners dynamically tune report structure, tone, and medical detail.
  2. May 2025 — Aug 2025

    Metal

    · AI Backend Engineer (Contract)

    Remote · Singapore

    • Built an intelligent network-discovery system analyzing LinkedIn/Gmail connections to find intro-paths to investors, backed by complex SQL (15+ filters).
    • Built an AI-enhanced investor search engine using pgvector embeddings; optimized PostgreSQL queries to sub-second on 100K+ profiles (500ms → <50ms).
    • Led the SuperTokens authentication migration to the latest versions.
  3. Feb 2024 — Feb 2025

    Mosaik

    · AI Engineer

    Remote · United States

    • Fine-tuned open-source LLMs (Phi, Llama) for query classification with Llama Factory.
    • Built a RAG-based chatbot (FastAPI) over client property documents and a Text-to-SQL system using RAG + LangChain + GPT.
    • Built an end-to-end image auto-labeling pipeline with GroundingDINO and AutoDistill.
  4. Mar 2023 — Jan 2024

    Beam

    · ML / Backend Engineer

    Remote · Germany

    • Built semantic routers using GPT for intent classification and fine-tuned a BERT model for query classification.
    • Developed an AI backend with LangChain, RAG, and FastAPI; designed a RAG pipeline for dynamic tool identification.
    • Engineered agent workflows via N8N connecting Slack, Gmail, and Google Docs; set up GitHub Actions for automated AWS deployment.
  5. Jan 2023 — Mar 2023

    Liquid Technologies

    · ML Engineer
    • Computer-vision model training with YOLOv7/YOLOv5, Roboflow, and image-processing techniques.
  6. Sep 2022 — Dec 2022

    University of Karachi

    · Research Assistant
    • Worked with a professor on deep neural networks for medical image segmentation, improving MRI analysis.