Haziq Ilham

A self-maintained profile · Kuala Lumpur, Malaysia

Featured engineerLast refreshed 2026

Muhammad Haziq Ilham (also known as Haziq Ilham) is a Malaysian software and machine-learning engineer1 based in Kuala Lumpur, Malaysia. He currently works at Rosary Labs4, where he builds production vision-language-model (VLM) inference pipelines, object-detection systems, and structured evaluation frameworks for real-world engineering-document understanding.

Ilham is known for shipping the full stack of an applied-AI product — from model to API to interface — and for an evidence-first habit of benchmarking before believing: a refactor of a core matching engine delivered a 160× speedup2, and his reinforcement-learning research on surprise-modulated learning rates improved on a PPO baseline by +14.8%.2 He has used the Anthropic Claude API and Claude Code SDK in production.6

Outside his day job, Ilham is a prolific builder whose public work ranges well beyond machine learning. His repositories include a from-scratch distributed key-value store in Go14, a cloud-agnostic parking “super-app” built as Go microservices on a hexagonal architecture15, and a piece of developer tooling — a GitHub App that gates pull-request merges behind LLM-generated comprehension checks.16 He works across Python, Go, Rust and TypeScript, and his side projects span reinforcement-learning research, MLOps, applied NLP, and hackathon builds — a curiosity-first approach that shows up everywhere in his work.

He graduated from Universiti Teknologi MARA (UiTM) with a Bachelor of Information Systems (Hons.) in Intelligent System Engineering, CGPA 3.64. He received the Vice Chancellor's Award and was on the Dean's List for six consecutive semesters.5

Career and experience

Ilham's professional work centres on applied AI for engineering documents at Rosary Labs, a software company building tools that turn technical drawings into structured, auditable data.4 He joined as a data-science intern in 2025 and transitioned to a full-time engineering role in 2026.

Junior Full-Stack Software Engineer — Rosary Labs

Jan 2026 — Present

Shipping a commercial AI product for engineering-document understanding: VLM inference pipelines, structured evaluation, and the systems around them.

  • Architected a structured evaluation framework decomposing model performance into Retrieval, Extraction, and Computation metrics for systematic error analysis.
  • Refactored the AI inference pipeline from parallel to a chained Object-Detection → VLM architecture with deterministic bounding-box-to-text alignment via Django ORM.
  • Benchmarked Table Transformer (TATR), CascadeTabNet & TableNet — 82.1s full-document detection on 92-page inputs — and recommended TATR on a performance–cost tradeoff.
  • Designed an AI-assisted BOQ population workflow (empty-field detection, fuzzy matching, cross-sheet extraction); evaluated Claude Code SDK vs Gemini for out-of-Excel computation.
  • Consolidated VLM calls into single batched API requests per grid image, cutting overhead while preserving coordinate-tag associations for auditability.
  • Migrated model deployment from Docker Hub to Azure Container Registry with secure auth for RunPod, eliminating external exposure risks in production.
PythonDjangoVLMClaude SDKRunPodAzure ACRDocker

Data Science Intern — Rosary Labs

Sep 2025 — Dec 2025

Built the detection-assisted extraction foundation for large-scale engineering drawings and squeezed orders-of-magnitude speedups out of the core engine.

  • Rewrote the core pattern-matching engine for a 160× speedup via caching, Pandas→NumPy migration, and large-scale synthetic event-pair generation over minute-level data (2022–2025).
  • Designed a full Object-Detection + VLM pipeline extracting engineering tags from large-scale P&ID diagrams, with YOLO / RF-DETR as a robust text-localization preprocessing stage.
  • Built experimental frameworks comparing VLM-only vs detection-assisted extraction, demonstrating measurable gains and surfacing failure modes through systematic gap analysis.
  • Proposed RF-DETR as a standalone RunPod serverless service — efficient batching, reduced memory footprint, faster startup via lazy-loading.
  • Diagnosed and resolved Redis OOM bottlenecks by profiling memory-intensive functions, improving backend reliability and throughput.
  • Used STUMPY for time-series similarity search and anomaly detection to stabilize correlation-based sensor monitoring.
PythonYOLORF-DETRRunPodRedisSTUMPYPandas/NumPy

Notable works

Shipped products and research, kept as a ranked index — most impactful first. Open a thread to read the case study and the numbers behind it.

Selected worksranked by impact · 6 entries
  1. 1.
    PDF-to-Excel Extraction Engine (rosary-labs.prod)
    196 points by exhazordinary 11 months ago | PRODUCTION · Rosary Labs |
  2. 2.
    Object-Detection → VLM Pipeline (rosary-labs.prod)
    280 points by exhazordinary 7 months ago | APPLIED ML · Rosary Labs |
  3. 3.
    Cognitive Temporal Reinforcement Learning (github.com/exhazordinary)
    166 points by exhazordinary 5 months ago | RESEARCH |
  4. 4.
    Curiosity-Driven Exploration Framework (github.com/exhazordinary)
    210 points by exhazordinary 6 months ago | RESEARCH |
  5. 5.
    Smart Street Object Detection (AIoT) (uitm.edu.my)
    335 points by exhazordinary 1 months ago | FINAL-YEAR PROJECT |
  6. 6.
    Cuckoo-Search Tower Allocation (github.com/exhazordinary)
    323 points by exhazordinary 8 months ago | OPTIMIZATION |
More → github.com/exhazordinary

Technical skills

HI
Haziq Ilham
Software & ML/AI Engineer · Kuala Lumpur, Malaysia
57
tags
● 6● 9● 14
Top tags
  • vision-language-modelsproduction VLM pipelines
  • object-detectionYOLO · RF-DETR
  • reinforcement-learningPPO research, +14.8%
  • pythonprimary language
  • pytorchdeep learning
  • claude-apiproduction LLM use
Languages
PythonTypeScriptJavaScript (ES6+)SQLC++RBash
Machine Learning
Vision-Language Models (VLMs)Object Detection (YOLO, RF-DETR)Reinforcement Learning (DQN, PPO, PPO-LSTM)Supervised & Unsupervised LearningTime-Series Forecasting & Anomaly DetectionModel Evaluation & BenchmarkingTransfer Learning · LoRA / PEFT
Deep Learning
PyTorchTensorFlowHuggingFace TransformersCNNs & LSTMsComputer VisionMeta SAM (Segmentation)Stable-Baselines3
LLMs & Agents
Anthropic Claude APIClaude Code SDKRAG (Retrieval-Augmented Generation)LangChainAutoGenOllama (local inference)Embeddings & Text Classification
MLOps & Infrastructure
DockerRunPod (serverless GPU)Azure Container RegistryFastAPICI/CD PipelinesSentry · OpenTelemetrypytestGit / GitHub
Backend
Django / Django ORMFlaskREST APIsWebSocketRedis (caching)SupabaseMongoDB
Frontend & Apps
React.js / Next.jsTailwind CSSStreamlitPlotly DashHTML5 & CSS3
Data
Pandas & NumPyPolarsETL PipelinesSTUMPY (matrix profile)Matplotlib & SeabornPower BI & Tableau
Cloud & IoT
AWS (S3, EC2)Azure (ACR)IoT / AIoT (ESP32, NodeMCU)
160× Engine speedup shipped3.64 CGPA · Dean's List ×620+ RL exploration methods92pg Docs parsed in production

Education

Ilham studied at Universiti Teknologi MARA (UiTM) in Shah Alam, Selangor, completing a Bachelor of Information Systems (Hons.) Intelligent System Engineering between Oct 2022 — Feb 2026.5 He graduated with a CGPA of 3.64 and earned a place on the Dean's List for six consecutive semesters.

  • Vice Chancellor's Award — top graduand recognition
  • Dean's List — 6 consecutive semesters
  • Head of Entrepreneurial Exco, AI Society (AIS) UiTM
  • Final-Year Project: Smart Street Object Detection (AIoT)

Certifications

  • Microsoft Certified: Azure Data Fundamentals
  • Asia Pacific University: Data Analysis with Power BI

Posts

The talk page. Thoughts from Haziq, and a space for anyone passing through to leave a note — recruiters, collaborators, fellow builders.

Pinned
HI
Haziq Ilhamauthor@exhazordinary · 2mo

I build production AI systems — VLM inference pipelines, object detection, and RL research. I benchmark before I believe, profile before I optimise, and measure before I claim. Open to SWE / ML roles. 🦾

💬 0🔁 14♥ 148
?
280
Loading…

In popular culture

An account of the events, as later recounted on an imageboard:

Anonymous shipped something Haziq No.160000
> be me, final-year UiTM student
> intern at an AI lab, told to make the matching engine faster
> profile it, cache it, swap Pandas for NumPy
> 160x speedup
mfw the benchmark finishes before the loading spinner renders
> promoted to full-stack engineer 3 months later
anon... just ship it

Notes

  1. Born and raised in Terengganu on Malaysia's east coast; now based in Kuala Lumpur.
  2. Received the Vice Chancellor's Award and made the Dean's List for six consecutive semesters.
  3. Served as Head of the Entrepreneurial Exco at the AI Society (AIS), UiTM.
  4. Works the whole stack of an applied-AI product — model, API and interface — and ships to production.
  5. Operating principle: benchmark before believing, profile before optimising, measure before claiming.
  6. Has shipped systems built on the Anthropic Claude API and the Claude Code SDK.
  7. Builds across Python, Go, Rust and TypeScript — from RL research to distributed systems and web apps.
  8. A serial side-project builder: his GitHub spans distributed systems, developer tooling, MLOps and hackathon entries.

References

  1. ^ 1.Aziz Azhar, Muhammad Haziq Ilham. "Curriculum Vitae" (PDF). Self-published
  2. ^ 2."exhazordinary — Repositories". GitHub. github.com
  3. ^ 3."Muhammad Haziq Ilham — Professional profile". LinkedIn. linkedin.com
  4. ^ 4."Rosary Labs — AI for engineering-document understanding". Employer
  5. ^ 5."Universiti Teknologi MARA — Faculty of Computer & Mathematical Sciences". uitm.edu.my
  6. ^ 6."Building with the Claude Developer Platform & Claude Code SDK". docs.anthropic.com
  7. ^ 7.Aziz Azhar, M. H. I. "Cognitive Temporal Reinforcement Learning: surprise-modulated learning rates improve PPO by +14.8% on LunarLander" (research project). github.com
  8. ^ 8.Aziz Azhar, M. H. I. "Curiosity-Driven Exploration Framework — 20+ intrinsic-motivation methods (RND, ICM, DRND, NGU, Go-Explore, DIAYN)" (research project). github.com
  9. ^ 9.Aziz Azhar, M. H. I. "Smart Street Object Detection (AIoT): YOLOv8 on ESP32-CAM" (Final-Year Project, UiTM, 2026). github.com
  10. ^ 10."Detection-assisted extraction of engineering tags from P&ID diagrams" (Object-Detection → VLM pipeline, Rosary Labs). Employer
  11. ^ 11."Microsoft Certified: Azure Data Fundamentals (DP-900)". learn.microsoft.com
  12. ^ 12."Data Analysis with Power BI — Asia Pacific University (APU)". apu.edu.my
  13. ^ 13.Raffin, A. et al. "Stable-Baselines3" — the RL library used for the PPO/PPO-LSTM experiments above. readthedocs.io
  14. ^ 14.Aziz Azhar, M. H. I. "distkv — a distributed key-value store" (Go). github.com
  15. ^ 15.Aziz Azhar, M. H. I. "Parking super-app — cloud-agnostic Go microservices on a hexagonal architecture, with wallet payments and multi-provider integration". github.com
  16. ^ 16.Aziz Azhar, M. H. I. "pr-comprehension-gate — a GitHub App that quizzes reviewers with LLM-generated comprehension questions before allowing a merge". github.com

External links