Nashit Bin Mashkoor

Full Stack and Machine Learning Engineer specializing in AI product deployment and MLOps.

Based in: Lahore, PakistanCurrently: Full Stack and ML Engineer, Independent ConsultingExperience: 6 years

Machine learning engineer with extensive experience building end-to-end AI products, from data collection and model fine-tuning to full-stack deployment. He combines expertise in Python, React, and MLOps to automate complex workflows and optimize real-time voice and vision applications.

Experience

Full Stack and ML Engineer

Independent ConsultingRemotefreelanceSep 2023 – Present

Summary

ADNP client data platform: Built and deployed a React application with FastAPI, PostgreSQL and background workers to normalize inconsistent financial spreadsheets, reducing processing time from 30 minutes to about 5 minutes. Worked directly with the client to review outputs against real datasets, investigate mapping errors and explain capacity tradeoffs; managed application deployment and client access. Voice applications: Built Python voice call workflows combining speech services, LLM tool calling and telephony, with CRM actions, call recording, voicemail handling and voice activity detection.

What I did

Built the frontend for a voice AI agent platform for a US startup.
Built a voice platform using LiveKit infrastructure with a config-based setup to swap TTS, STT, LLM, and telephony providers based on cost and client requirements.

Results

Reduced processing time from 30 minutes to about 5 minutes for financial spreadsheet normalization.

ReactFastAPIPostgreSQLPythonLLMCRMVoice Activity DetectionTypeScriptCartesiaDeepgramOpenAI

Machine Learning Engineer

Noze.caRemoteworkAug 2022 – Present

Summary

Automated ML training and deployment with Kubeflow, Kubernetes and TensorFlow Serving, reducing model retraining from 24 hours to 2 hours. Built FastAPI services to process and evaluate annotation data, increasing data yield by 30%; automated study and model specific dataset curation with Airflow, reducing manual effort by 95%. Implemented data quality validation, reducing data errors by 35%; made annotation datasets queryable through partitioned Parquet on AWS S3, Glue and Athena. Created a Python SDK for external teams integrating with data collection services, plus Retool administration tools and Grafana monitoring dashboards. Built a LangGraph application answering natural language questions across PostgreSQL, APIs and documents; added pytest regression checks for LLM responses before releases.

What I did

Wrote Python SDKs to facilitate platform integration across teams.
Took the initiative to create an intuitive SDK that enabled teams to manage IoT devices and experiments independently using their existing Jupyter notebooks.

Results

Enabled other teams to directly integrate the cloudscent platform into their workflows by writing Python SDKs, removing the need for custom workflow development.
Reduced model retraining from 24 hours to 2 hours.
Increased data yield by 30%.
Reduced manual effort by 95% through automated dataset curation.
Reduced data errors by 35%.

KubeflowKubernetesTensorFlow ServingFastAPIAirflowParquetAWS S3AWS GlueAWS AthenaPython SDKRetoolGrafanaLangGraphPostgreSQLpytestLLMpycloudscent

Machine Learning Engineer

Xavor CorporationLahore, PakistanworkAug 2020 – Jul 2022

Summary

Expanded a training dataset from 4,000 to 14,000 images and fine tuned YOLOv4, improving accuracy from 70% to 83% while increasing supported classes from 3 to 15. Reduced fall detection frame processing time by 93% through CNN/LSTM optimization and a faster pose estimator; deployed inference on NVIDIA Jetson Nano. Built AWS Kubeflow pipelines connecting data ingestion, model training and deployment; stored training data and model artifacts in S3.

Results