Muhammad Ihza Mahendra

Experience

Zero · Dubai, UAE

June 2025 - Present

Build, iterate and maintain multimodal AI agents (voice + vision + text) that simulate real-world workplace demonstrations, where users walk AI coworkers through the project work they have accomplished, hitting sub-500ms first-audio via LiveKit + Deepgram + ElevenLabs on top of Claude and Gemini.
Own, iterate and maintain the core user-facing flow along with the evaluation and calibration stack (LLM-as-judge grading, rubric design, real-time signal extraction) that build a deep, evolving understanding of each user and guide them toward effective, engaged learning.
Design, iterate and maintain the multi-agent orchestration behind the interactive scenarios: tool use and function calling, streaming voice, Computer Use over live screen state, and a coordinate-mapping formula so the AI can point at the right pixel on the user's screen for guiding purpose across any window size.
Build, iterate and maintain the LLM engineering and infra stack: multi-provider routing and failover (Claude, Gemini, OpenAI, Anthropic SDK, LiteLLM), Bedrock prompt caching with cost attribution and observability, Temporal workflow orchestration, and async task pipelines.
Deploy, manage and maintain production infrastructure on AWS Fargate, ECS, SageMaker and HyperPod with Docker + Kubernetes, CI/CD via GitHub Actions, Sentry for observability.

Bytedance · Jakarta, Indonesia

Jan 2024 - June 2025

Implemented the new recommendation engine on the TikTok Shop landing page, increasing GMV by up to 129%.
Developed additional lighter ranking modules aligned to business goals, ranking a wider product pool to pass on to the final ranking model and personalize user preference. Drove +13% GMV, +9.24% orders and +0.73% CTR.
Developed features and model architecture for ranking and recall modules in a recommendation system, using deep learning on ByteDance's cloud. Drove +93% GMV, +96% orders and +39% CTR.

Tokopedia · Jakarta, Indonesia

Dec 2021 - Dec 2023

Designed and implemented automated audience distribution in push notifications for user category targeting, using a deep-learning model that predicts a user's next category from sequence behaviour. Reached 0.80% CVR and 0.45% CTR across 150M+ targeted users.
Developed a Gen-AI use case for e-commerce by building an AI shopping assistant API endpoint using GPT on Azure, applying instruction tuning to keep the assistant helpful, safe, relevant and inspirational.
Built a LightGBM-based surrogate income model to predict Tokopedia credit card eligibility, achieving +0.75% TP uplift with minimal increase in false negatives.
Developed an end-to-end CTR prediction model for push notifications, boosting CTR by 153%, while managing the project lifecycle.

Skills

Programming Languages: Python, C++, TypeScript
Backend & Orchestration: FastAPI, Flask, Pydantic v2, asyncio, Temporal, WebSockets
LLM & Agent Tooling: Claude, Gemini, OpenAI, Anthropic SDK, LangChain, LangGraph, LiteLLM
Real-time & Voice: LiveKit, Deepgram, ElevenLabs
Classical ML: TensorFlow, LightGBM, Scikit-Learn, Pandas, NumPy, Matplotlib, Deep Learning, OpenCV, Pytesseract
Data & Storage: PostgreSQL, pgvector, MySQL, Hive SQL, Redis, Supabase, ChromaDB
MLOps & Infra: Docker, Kubernetes, AWS Fargate, AWS ECS, AWS SageMaker, AWS HyperPod, GCP, Airflow, GitHub Actions, Jenkins, Sentry, Bytecloud, Distributed Inference

Personal projects

AI-based Personality Test: Designed a speech-to-speech application for personality testing using an open-source LLM.
Verbal Trainer AI: Built the verbal trainer AI skeleton with an optimized LLM model.

Publications

Education

Telkom University · 2017 - 2021

GPA 3.81, Cum Laude