Skip to content
View pr1tzy's full-sized avatar

Highlights

  • Pro

Block or report pr1tzy

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
pr1tzy/README.md
Prithvi Singh — ML infrastructure and backend. Speed trace from a real Monza lap, 1:28.197 against a 1:25.750 ghost.

The trace above is a real lap: Monza, Lotus Exos 125 S1, recorded by Ghostline. Violet is my lap, cyan is the ghost.

I work on the systems side of machine learning. Distributed training, retrieval, and the infrastructure that keeps models serving after the notebook is closed.

training     multi-node GPU pretraining of dense embedding models
retrieval    RAG over Qdrant and Neo4j, masking strategy ablations
backend      FastAPI, PostgreSQL, Alembic, REST
infra        AWS (ECS, S3, SQS, Cognito), Terraform, Docker, Slurm
portfolio    prithvisingh.fyi
shipped      ghostline.prithvisingh.fyi

Selected work

Ghostline — 60 Hz lap telemetry, ghost comparison, corner-level time loss. 4.7K LOC, 4 test suites, 3 CI workflows, 4 releases. Live. Self-healing LLM QA — continuous eval against benchmark and live traffic, auto rollback. DVC, MLflow, Jenkins, 12 pipeline modules. ParkiSense — keystroke dynamics and drawing tremor, wavelet features. SVM and Random Forest, approx 82% validation.

Ghostline is live at ghostline.prithvisingh.fyi.


Currently

Backend engineer at AceLink Software Solutions, working on Python services and AWS infrastructure.

Before that, eight months as a technical intern at BharatGen (IIT Bombay), working on distributed pretraining of dense embedding models on a multi-node GPU cluster: masking strategy ablations through pretraining and SFT, Slurm cluster tooling, and large-scale data migration to S3.

Reading about inference optimisation and distributed training internals.


Stack

languages    Python, SQL, TypeScript, Lua
ml           PyTorch, Transformers, scikit-learn, XGBoost
training     Multi-node GPU, Slurm, Fabric, distributed data pipelines
retrieval    Qdrant, Neo4j, embeddings, RAG
mlops        DVC, MLflow, Jenkins, GitHub Actions
backend      FastAPI, SQLAlchemy, PostgreSQL, Alembic
infra        AWS, Terraform, Docker, Cloudflare

Research

Do UN Discussions Reduce Terrorism? An Empirical Study on Political Stability and Terrorism Casualties — B.Tech capstone, Department of Data Science, MPSTME.

Text-mined a UN Attention Score from UN Official Document System reports and merged it with Global Terrorism Database casualties and World Bank governance indicators into a 1995–2020 country-year panel. Estimated pooled OLS, fixed effects, random effects and two-step difference/System GMM to handle persistence and endogeneity. Political stability holds a robust negative association with casualties; discourse volume alone does not.


Activity over the last 30 days across all repositories. Language mix across public repositories, source code only, notebooks excluded.

prithvi30singh@gmail.com · +91 9619952004 · LinkedIn · prithvisingh.fyi

Pinned Loading

  1. ghostline ghostline Public

    Telemetry analysis for Assetto Corsa: 60 Hz lap recording, ghost comparison, and corner-by-corner time loss.

    Python 1

  2. ParkiSense ParkiSense Public

    Early Parkinson's screening from keystroke dynamics and drawing tremor, using wavelet and time-frequency features.

    Jupyter Notebook 1 1

  3. self-healing-llm-qa self-healing-llm-qa Public

    Self-healing LLM QA service: continuous evaluation against a benchmark and live traffic, with automatic rollback on regression.

    Python