Skip to content
Roshil Shrestha
Portrait of Roshil Shrestha

Roshil Shrestha
ML Engineer Associate, Fusemachines

I turn complex business data into ML systems that reach production.

ML Engineer Associate at Fusemachines since 2023. I build end to end: graph neural networks that explain their predictions, real-time detection and tracking, demand forecasting services and LLM agents that remember.

GitHub LinkedIn

Graph MLVisionLLM agentsNLPForecasting
Client project Proof of conceptPick a technique to filter the work below.
30 FPS
real-time ball and player tracking for Darkhorse Soccer Analytics
8
people led in the soccer annotation team
9
ML systems built: 3 for clients, 6 proofs of concept
3+ years
in ML, now ML Engineer Associate at Fusemachines

Why hire me

What I bring to a team, and the project that shows it.

  • Graph ML that explains itself

    GNNs and XGBoost on connected player data, with SHAP and Gemini turning every prediction into a plain reason.

    Player 360
  • Vision that runs in real time

    Detection and tracking tuned for small, fast objects in high-resolution video, at 30 FPS.

    Soccer tracking
  • Forecasts the business can use

    Item-level forecasts reconciled with quarterly plans and delivered as reports, straight from BigQuery.

    Demand forecasting
  • Agents that remember

    LLM agents with persistent memory, knowledge graphs and tools, tested across sessions.

    Agent memory
  • NLP on real documents

    Fine-tuned and zero-shot transformers that classify documents and pull out the entities that matter.

    Document verification

Work

Selected case studies: three client systems and six proofs of concept, built at Fusemachines. Most of this code isn’t public, so each one has a full write-up.

Showing all 9 projects

Read the case study

Entain, Mar – Aug 2025

Early VIP & High-Potential VIP Player 360

3–21 daysbehavioural windows for early VIP detection

An ML and graph analytics platform that spots future VIP players within their first 3 to 21 days, with a plain-language reason for every prediction.

  • Behavioural, monetary, temporal, engagement and cross-product features covering deposits, stakes, active days, sessions, NGR and product mix.
  • Graph features for connectivity, neighbourhood statistics, propagation signals, behavioural clustering and shared-entity patterns, with relationship filtering to cut graph noise.
Graph MLForecastingLLM agents
Read the case study

Sanmina

Demand Forecasting System

An API-driven forecasting service that produces item-level weekly forecasts from BigQuery or uploaded data and turns them into client-ready Excel reports.

  • Preprocessing that turns demand, MDS, capacity, WIP, inventory, CTB and calendar data into clean, continuous ITEM × WEEK time series.
Forecasting
Read the case study

Darkhorse Soccer Analytics

Soccer Detection & Tracking

30 FPSreal-time detection and tracking

Detects and tracks the ball, players and jersey numbers in match video at 30 FPS, backed by an annotation team I led.

  • SAHI (Slicing Aided Hyper Inference) with 6-slice and 8-slice configurations for small-object detection in high-resolution frames.
Vision
Read the case study

Tampa General Hospital

Patient Monitoring & Pose Estimation

Tracks patient movement and posture in real time with YOLOX, RTMO and ViTPose, and uses Florence to drive automated reports and alerts.

  • YOLOX for patient detection and RTMO for real-time multi-person pose estimation.
VisionProof of concept
Read the case study

Proof of concept

Persistent Cross-Session Memory for AI Agents

Gives Google ADK agents a PostgreSQL-backed memory, so they recall past sessions, preferences and instructions.

  • Evaluation questions covering temporal reasoning, open-domain knowledge, adversarial cases, and multi-hop and single-hop QA.
LLM agentsGraph ML
Read the case study

Proof of concept

Sales Copilot

Agents that create and look up HubSpot deals, a Neo4j graph of the CRM, and natural-language answers with next-step recommendations.

  • A knowledge graph of deal stages, deals, companies, notes, contacts and emails for contextual querying.
LLM agentsGraph ML
Read the case study

Proof of concept, Sep – Oct 2025

Fashion Mood Boards with Gemini

Generates editorial fashion visuals and mood boards from prompts and reference images, so designers can try a concept without a photoshoot.

  • Multiple reference images combined with prompts to produce collages and mood boards.
LLM agentsVision
Read the case study

Proof of concept

Project Management System Agent

Merges data from multiple Monday.com boards into cross-board summaries, recommendations and email drafts.

  • Cross-board analysis for holistic project summaries and recommendations.
LLM agents
Read the case study

A Streamlit app that classifies uploaded documents, predicts dispute codes from comments and extracts entities.

  • Fine-tuned RoBERTa sequence classification that predicts dispute codes from user comments.
  • Zero-shot BART classification of document types, without training a model per type.
NLP

Experience

  1. 2018

    Higher Secondary (+2) in Science

    Shikshadeep HSS, HSEB

  2. 2023

    BSc CSIT

    AIMS College, Tribhuvan University

  3. Aug 2023

    ML Engineer Associate Trainee

    Fusemachines

  4. Nov 2023 – now

    ML Engineer Associate

    Fusemachines

  • Nov 2023 – Present

    Fusemachines

    ML Engineer Associate

    • Built a Player 360 platform for Entain that flags Early and High-Potential VIP players using BigQuery, XGBoost, homogeneous and heterogeneous GNNs, SHAP and Gemini-generated explanations. Case study
    • Built an API forecasting pipeline for Sanmina with hierarchical MLForecast (XGBoost) models and automated Excel reporting on supply, inventory and capacity. Case study
    • Developed computer vision pipelines to detect and track players, balls and jersey numbers for Darkhorse Soccer Analytics, and led an 8-member annotation team. Case study
    • Developed a real-time patient monitoring and pose estimation system for Tampa General Hospital with YOLOX, RTMO, ViTPose and Florence. Case study
    • Built LLM agent proofs of concept: a HubSpot sales copilot on Neo4j knowledge graphs, a Monday.com project management agent, and persistent memory for Google ADK agents. Case study
    • Built a document verification and classification app with fine-tuned RoBERTa, zero-shot BART and GLiNER. Case study
  • Aug 2023 – Nov 2023

    Fusemachines

    ML Engineer Associate Trainee

    • Joined the Computer Vision Upskilling programme and built a foundation in computer vision techniques with Python, OpenCV and Matplotlib.
    • Learned AWS services including S3, EC2 and Lambda.
    • Applied Python naming and code standards, contributed to code cleaning and optimisation, and studied error analysis techniques.
    • Annotated data for Fuse Extract and Darkhorse.

Education

  • 2023

    BSc in Computer Science and Information Technology (BSc CSIT)

    Alliance for Innovative Management and Social Studies (AIMS) College, Biratnagar

    Tribhuvan University

  • 2018

    Higher Secondary (+2) in Science

    Shikshadeep Higher Secondary School of Science

    HSEB

Toolbox

The languages, models, frameworks and platforms I work with, grouped by what they’re for.

Languages

  • Python
  • C/C++
  • SQL

Machine learning

  • scikit-learn
  • XGBoost
  • TensorFlow
  • Keras
  • PyTorch
  • SHAP
  • Transfer learning
  • NumPy
  • Pandas
  • Matplotlib
  • Seaborn

Graph ML

  • PyTorch Geometric
  • Homogeneous & heterogeneous GNNs
  • GAT
  • Neo4j
  • Graphiti

Computer vision

  • YOLOv8
  • YOLOX
  • RT-DETR
  • SAHI
  • RTMO
  • ViTPose
  • Florence
  • OpenCV
  • Pillow
  • Kalman filters
  • Roboflow
  • CVAT

NLP

  • Hugging Face Transformers
  • BERT
  • RoBERTa
  • BART
  • GLiNER
  • spaCy
  • NLTK
  • Named entity recognition
  • Text classification
  • Summarisation
  • Sentiment analysis

LLMs & agents

  • Google ADK
  • LangChain
  • CrewAI
  • Gemini
  • OpenAI
  • RAG
  • CopilotKit
  • Prompt engineering
  • Multimodal image generation

Forecasting

  • MLForecast
  • Hierarchical forecasting
  • Time-series feature engineering

Data

  • Google BigQuery
  • PostgreSQL
  • MySQL
  • MongoDB

Cloud & MLOps

  • Google Cloud
  • AWS (EC2, S3, Lambda)
  • Microsoft Azure
  • Docker
  • API model serving
  • GPU computing

APIs & apps

  • FastAPI
  • Flask
  • Streamlit
  • REST APIs
  • HubSpot API

Tools

  • Git
  • Jupyter
  • VS Code
  • Google Colab

Projects

Open-source projects on GitHub, code included. My client work stays private, so its write-ups are under Work.

github.com/Roshil

Contact

Questions about my work, a role, or a project you’d like help with? Send me a note and I’ll get back to you.

Opens your email app with this message filled in.