Haadhi Irfan

AI Enablement Specialist · Imobisoft

Sole builder of the branch's internal tooling, shipping full-stack products solo, concept to live. Field-tested an orchestrated multi-agent coding workflow by building BulkUp (Expo/React Native + Supabase, offline-first sync), writing the measured findings into a reusable delivery methodology. Built a photo-to-macro scanner with a server-side Llama-4 vision proxy, and agentic developer tooling (TheBench, TheCompanion) adopted across a 16-engineer, 5-team org.

• React Native
• TypeScript
• Supabase
• Llama-4 Vision
• Agentic AI

AI Engineering Consultant · Aspect

Architected Navigator Copilot, a production-grade multi-agent LLM orchestration microservice, using LangGraph, LangChain, FastAPI, and Groq (Llama-3.3-70b). Enabled natural language querying of engineer KPI data across 18 performance metrics and 6 scoring categories. Built a 3-pass fuzzy name resolution algorithm achieving 93% intent-routing accuracy, async data pipelines with L2 caching, and delivered 210+ automated tests with 100% pass rate.

• LangGraph
• LangChain
• FastAPI
• Groq
• React
• Python

Full-Stack Software Engineer · Einstein11Plus

Designed and implemented a webhook-driven booking and automation engine integrating Shopify events with internal MySQL systems. Automated student enrollment across 4+ product workflows, reducing manual operations by ~80%. Built frontend components with MutationObservers that cut response times by ~30x (from 6s to <200ms).

• JavaScript
• PHP
• MySQL
• Shopify API
• Webhooks

AI Consultant · Muslim Hands (NGO)

Advised senior executives (Deputy CEO & Senior Management Team) on systemic AI adoption, identifying 3+ high-impact use cases that streamlined planning and mitigated operational risks. Developed a secure GDPR-compliant AI data strategy and designed 3+ interactive corporate training workshops on prompt engineering and LLM tools.

• LLMs
• Prompt Engineering
• GDPR
• AI Strategy

Deep Learning Intern · University of Manchester

Researched time-series-to-graph representation learning for long-horizon forecasting. Built a custom Deep Visibility Series algorithm with a CNN backbone, integrated multi-head self-attention into LSTM baselines (~3% error reduction), and ported training to the CSF4 HPC cluster with parallel SLURM jobs, a 10x compute speedup.

• PyTorch
• CNN
• Self-Attention
• HPC / SLURM

NOC / Deep Learning Intern · Sterlite Technologies

Integrated an LLM-powered self-service conversational assistant into an enterprise NMS serving 100+ users. Conducted pretraining over 200M samples and SFT over 10K prompt-response pairs. Engineered an RLHF alignment pipeline with reward modeling on 30K preference pairs, slashing mean time to resolution by ~35%.

• PyTorch
• RLHF
• Reward Modeling
• Bash
• Linux

MizanBench: Open Benchmark for LLM Reasoning on Islamic Finance

Independent · open-source release at v1

The first systematic evaluation of how well language models reason about Sharia-compliant finance, with structured cases across riba, murabaha, zakat, ijara, gharar, and sukuk, each with gradeable key points derived from AAOIFI standards. A two-pass harness (candidate answers cold, an LLM judge scores against the reference) produces per-model verdicts and a leaderboard. v1 ships with 100+ cases and frontier-model results; the repository goes public with it.

• LLM Evaluation
• Benchmark Design
• Islamic Finance
• Python
• Claude API

Navigator Copilot: Multi-Agent KPI Microservice

Aspect · shipped to production

Production multi-agent LLM orchestration microservice for natural-language querying of engineer KPI data across 18 metrics and 6 scoring categories. LLM-powered intent routing with 3-pass fuzzy entity resolution reached 93% accuracy on a QA benchmark; delivered with 210+ automated tests at a 100% live pass rate, role-based leaderboards behind enforced access tiers, and namespaced multi-tier caching.

• LangGraph
• LangChain
• FastAPI
• Groq
• Python

StAnify: Multi-Agent Visual Storytelling & AI Tutoring

University of Birmingham · Advisor: Dr. Mohammad Bahja

Designed a distributed multi-agent pedagogical framework using CrewAI for early childhood education (ages 5–9). Integrated a Qdrant RAG pipeline for zero-hallucination content. Blinded A/B testing with 30 evaluators showed StAnify was preferred in 65% of cases over GPT-4, with +1.59 in visual clarity and +1.51 in engagement. Compressed lesson prep from 50–80 mins to ~3 mins (17–25x speedup) at ~$0.15 per package.

• CrewAI
• Qdrant
• Stable Diffusion XL
• Pydantic
• Streamlit

DVS-LSTM: Novel Hybrid Time-Series Architecture

IIT Kharagpur + University of Manchester

Co-designed a custom hybrid architecture combining visibility graph mappings with multi-head self-attention LSTM for long-horizon climate forecasting over 50+ years of data. Outperformed all deep learning baselines by 2–5% on MSE/MAE. Achieved 10x compute speedup by porting from local notebook to HPC cluster with parallelized SLURM configurations (~32 cores).

• PyTorch
• CNN
• LSTM
• Attention
• HPC / SLURM

AI Birmingham · Fautons

Co-speaker · Fautons is an official Anthropic Claude Partner Network member

Weekly hands-on sessions teaching practical AI-agent workflows: how tools like Claude Code deliver real leverage for business owners, students, and professionals, not just developers. Sessions blend fundamentals with advanced techniques and are built around skills attendees can use straight away.

• AI Agents
• Claude Code
• Practical Workflows
• AI Adoption