Prakash Chaudhary

---

Summary

Machine Learning Engineer with over 5 years of experience designing and deploying scalable end-to-end ML models, with expertise in building AI agents leveraging human-in-the-loop automation to drive measurable business impact.

---

Industry Experience

Fetch | Remote (AI Labs)

• Tech Lead, Agentic Infrastructure: Lead the infrastructure work behind production agents, from evaluation and deployment to observability and day-to-day operation
• Built and shipped AI Shopping Assistant serving a 200K-user experiment -- a Python/LangGraph agent with product card integration with shoppable product cards (catalog search, web search, 4.2M normalized images at 18x cost efficiency), episode-based conversation history (DynamoDB + S3), and production observability (Opik, OpenTelemetry, Grafana), deployed on AWS ECS Fargate
• Built end-to-end LLM evaluation system -- a continuous improvement loop connecting offline evaluation (golden dataset), online monitoring (production traces), and human annotation (judge calibration), with CI/CD deploy gating that blocks releases on quality regression
• Led Multi-Agent Matching Framework: cross-functional team of 4 engineers; improved match rates by 3.1-3.3 percentage points (73.8%→76.9%) unlocking $165M in GMV; achieved 15% full automation with LLM-as-a-Judge guardrails at 97% accuracy, automating 50+ hours of manual operations to 2 minutes
• Developed 3-agent annotation workflow using LangGraph with LLM annotation at 80%+ accuracy to reduce double-annotation to single annotator, cutting total annotation workload by 40% while maintaining dataset quality through human tie-breaking on disagreements
• Fully automated RAG pipeline to flag incorrect product matches at 95% recall, protecting ~$800K in GMV annually; replaced manual workflow, reducing daily effort from 8 hours to 15 minutes and saving 40 hours weekly

Fusemachines Nepal | Kathmandu, Nepal

• Developed multimodal contrastive learning pipeline combining image and time-series data for sales forecasting of new products without historical data
• Led efforts to build an ensemble model to classify fashion products into Kate Spade's four customer segments using computer vision and scipy optimize, achieving 62% accuracy
• Developed an in-house recommendation system using Factorization Machine model, AWS EC2 and Lambda, achieving ranking metrics comparable to AWS's recommendation engine
• Implemented Multi-Quantile RNN-based sales forecasting model for Coach fashion, reducing stock outs by 30%

---

Academic Experience

NASA-IMPACT | Huntsville, AL

• Developed a hierarchical document classification baseline model for Earth Science, achieving 46% exact match accuracy across three hierarchical levels, enhancing document organization and retrieval

Fusemachines Nepal | Kathmandu, Nepal

• Taught 50+ students in Data Science, covering Python, Git, REST API, SQL, LLM, and MLOps as part of the AI Fellowship Latin America 2023
• Taught Computer Vision as a lecturer and assisted with Machine Learning and Deep Learning courses as a Teaching Assistant for the AI Fellowship Nepal in 2022

---

Education

University of Alabama in Huntsville | Huntsville, AL

**GPA:** 4.0

• Relevant Courses: Deep Learning, Survey AI, Algorithm, Big Data Computing, Software Engineering
• Awards: Phi Kappa Phi Honor Society

MITx on edX | Online

**Grade:** B

• Relevant Courses: Probability, Statistics, Machine Learning, Data Analysis

---

Projects

Lazyflow - Calendar-First Todo App Dec 2025 - Present

*A free, open-source iOS app that bridges task management with calendar scheduling, featuring on-device AI.*

• Built and published a native SwiftUI iOS app on App Store with full Apple ecosystem support (Watch, Widgets, Dynamic Island, Siri Shortcuts) and on-device Apple Intelligence for smart task suggestions while maintaining privacy with local-first data storage

diskard - Developer-Aware Disk Cleanup Feb 2026 - Present

*An open-source Rust CLI/TUI tool that scans for reclaimable disk space from build caches, AI models, and dev tools.*

• Built a parallel scanner with 18 recognizers and interactive ratatui TUI with drill-down inspection, disk visualization, and safe trash-based cleanup; published on crates.io and Homebrew

Spectral Unmixing using Machine Learning Mar 2023 - Oct 2023

*A ML system for material characterization using spectral signal from Infrared (IR) spectroscopy.*

• Developed an ML pipeline, including signal processing, feature engineering, and a multi-label neural network model, achieving over 95% F1 score for predicting material composition from spectral signatures across diverse datasets

---

Technical Skills

LLM & Agentic AI: LangGraph, LangChain, RAG, Multi-Agent Systems, Fine-Tuning (LoRA), RLHF, Prompt Engineering (Chain-of-Thought, ReAct), LLM-as-a-Judge Evaluation, Embeddings & Vector Search

ML/AI: Computer Vision, NLP, Time Series Forecasting, Recommendation Systems, Statistical Modeling

Frameworks & Tools: PyTorch, TensorFlow, Hugging Face Transformers, Distributed Training (DDP), Docker, Kubernetes

LLM Ops: Opik, Langfuse, DeepEval, OpenTelemetry, Model Serving (TorchServe, Ollama), Quantization

Infrastructure: AWS (SageMaker, EC2, Lambda, S3, DynamoDB), Apache Spark, Kafka, Snowflake, Vector DBs (FAISS, Pinecone, Chroma), FastAPI, MCP

Programming: Python, Rust, Go, C++, SQL