abdul hadi

AI Engineer specializing in multi-agent orchestration and high-performance model optimization.

Based in: Karachi, PakistanCurrently: Founder & Engineer, Sodeom

An experienced AI engineer and founder with 3 years of expertise in building end-to-end LLM systems, multi-agent orchestrators, and FastAPI backends. He specializes in model quantization and distillation research to deploy high-performing small models on constrained hardware.

Experience

Founder & Engineer

SodeomRemotefounderPresent

Summary

Built and operate a privacy-first search engine, evolving it from direct result-page crawling to a SearXNG-based meta-search backend.

What I did

Built Aura Counter, a Sodeom side project with a full account system (signup, login/logout, sessions, cookies, email).
Maintain Flask-based services behind Cloudflare, with an agent monitoring uptime and operations.
Transitioned from direct crawling to a SearXNG-based meta-search backend to reduce maintenance costs associated with individual engine scrapers.
Managed search latency and ranking by configuring specific engine selections and timeouts within the SearXNG framework.

SearXNGFlaskCloudflare

Founder & AI Engineer

TrillionirAIRemotefounderPresent

Summary

Built and run a flat-rate LLM inference API with OpenAI- and Anthropic-compatible routing, so existing SDKs and tools work with no code changes.

What I did

Built a model evaluation pipeline on openbench with a custom OpenAI-compatible proxy, benchmarking Gemma 4 31B, Devstral 24B, Rnj-1 8B and SmolLM 135M.
Designed the product around reliability: an API that always responds instead of failing with out-of-credits errors.
Managed multi-provider routing to handle provider failures, rate limits, and fallbacks to prevent out-of-credits errors.
Tested speculative decoding using llama.cpp on company servers, though it was not implemented due to hardware performance limitations.

Results

Served 21 users
Processed approximately 371,000 requests over a few months
Averaged roughly 3,000 to 4,000 requests per day

OpenAI SDKAnthropic SDKopenbenchGemma 4 31BDevstral 24BRnj-1 8BSmolLM 135M

Researcher

TriMind (open source)volunteerPresent

Summary

Compared FP32, INT4, INT2, Q2_K and BitNet precision tiers on Qwen3-8B using Kaggle 2×T4 GPUs, LLaMA-Factory and TRL.

What I did

Published a quantized model, thetrillioniar/Qwen3-8B-1Q, on HuggingFace and shared findings publicly.
Fine-tuned Kimi-VL-A3B-Thinking with QLoRA and DeepSpeed ZeRO-2 on 2×T4.
Working on teacher-student distillation of Qwen3.6-35B-A3B into Falcon 3B for coding and agentic tasks, including distillation dataset creation.

Results

Benchmarked Qwen3-8B precision tiers from FP16 down to 1-bit using MMLU to evaluate knowledge-accuracy trade-offs.
Reduced model memory footprint from 16 GB at FP16 to approximately 1.1 GB at 1-bit precision.

Kaggle 2×T4 GPUsLLaMA-FactoryTRLQLoRADeepSpeed ZeRO-2Qwen3-8BKimi-VL-A3B-ThinkingFalcon 3BopenbenchMMLU benchmarkingModel quantization analysis

Skills

Technical

Python

Multi-agent orchestration

FastAPI

Ollama

REST APIs

QLoRA

Flask

MCP

n8n

GGUF/llama.cpp

Cloudflare

Kaggle GPUs

SearXNG

Node/TypeScript

DeepSpeed ZeRO-2

Ubuntu server administration

Intent routing

Deterministic policy enforcement

MMLU benchmarking

Model quantization analysis

Speculative Decoding

Projects

TriageDesk

Developerpersonal

Multi-agent customer support orchestrator. Python/FastAPI service with a custom BaseAgent orchestrator built directly on the OpenAI SDK (no heavy framework). Previously ran on an Ubuntu VM exposed via Cloudflare Tunnel.

PythonFastAPIOpenAI SDKUbuntu VMCloudflare Tunnel

AI blog and short-form video pipeline

Developerpersonal

Cron-driven pipeline that writes posts for blogs.thetrillioniar.me using MCP tool use for web search, and turns them into daily short-form videos that auto-upload across platforms. Evaluated and selected models by tool-calling reliability and quota efficiency.

CronMCPWeb Search Tools

Visit the site

Automation and outreach agents

Developerpersonal

An n8n workflow that posts daily LinkedIn updates from GitHub commits through the official API, and a multi-step outreach agent that finds founder contacts, writes lead CSVs and sends emails via Resend, running on Gemma 4 31B through Ollama.

n8nLinkedIn APIGitHub APIResendGemma 4 31BOllama