Prithvi Singh
The show
Prithvi Singh, smiling, in an orange kurta
Prithvi Singh
as himself

Call sheet

Prithvi
Singh

Backend and machine learning engineer

Available
Immediately
Works
Remote, hybrid or on-site
Relocation
Open to anywhere
Based in
Mumbai

Contacts

Résumé, PDF

I build backend systems and train models. Right now I write the APIs and cloud infrastructure behind a venue platform at AceLink. Before that I trained a 2.9-billion-parameter embedding model on H100 clusters at BharatGen, IIT Bombay.

Looking for

A team that trains AI models or runs them in production every day. I want to work where models are the product, not an afterthought: pretraining and fine-tuning, retrieval and embeddings, evaluation, or the backend and infrastructure that serves them to real users. ML engineer or backend engineer on an AI team; I can start immediately.

Parameters in the embedding model I trained
2.9B
Data moved to S3 with a tool I built
1 PB+
Validation accuracy, Parkinson’s screening from typing
82%

Work

II

Backend Developer

  • Build and maintain REST APIs in Python with FastAPI and SQLAlchemy: routers, request validation and the business rules behind them.
  • Design the PostgreSQL schema and ship it through Alembic migrations as the product grows.
  • Provision the AWS side with Terraform: containerised services on ECS, identity through Cognito.

Python, FastAPI, PostgreSQL, Alembic, AWS, Terraform

AceLink Software SolutionsMumbai

Jul 2026 to now

I

Technical Intern

  • Trained a 2.9B-parameter dense embedding model on a distributed H100 cluster from a tokenised English corpus, cutting inference latency and compute overhead.
  • Compared causal and bidirectional masking in pre-training and fine-tuning, which reduced hallucinations and improved retrieval accuracy.
  • Built a full-stack tool (FastAPI, React) to migrate more than a petabyte to S3, and automated GPU diagnostics with Fabric on Slurm.
  • Developed retrieval-augmented applications that combine Qdrant vector search with Neo4j graph retrieval.

PyTorch, distributed training, Slurm, RAG, Qdrant, Neo4j

BharatGen, IIT BombayMumbai

Oct 2025 to May 2026

Projects

III

Forex forecasting and optimisation suite

An LSTM pipeline for short-term currency moves that detects the market regime (bull, bear, sideways) and adapts its strategy to it, with Markowitz portfolio optimisation, arbitrage detection and backtesting on historical data.

LSTM, portfolio theory, regime detection, backtesting

IV

Parkinson’s diagnosis toolkit

Screens for early Parkinson’s signs from how people type and draw. Wavelet and time-frequency features feed SVM and random-forest models that reach about 82% validation accuracy, with dashboards comparing patients against healthy controls.

Signal processing, keystroke dynamics, scikit-learn

V

Mumbai temperature prediction

Twelve years of monthly climate data modelled three ways, ARIMA, SARIMA and XGBoost, to compare classical statistics with machine learning on long-term temperature trends in the city.

Time series, ARIMA, SARIMA, XGBoost

Now

Reading, September 2026. Five papers from the last three months, and what each one changes.

  1. NCP-ArchPreviewPretraining on concepts that span several tokens, not just the next token. It reaches OLMo-3-7B’s loss on about half the data.
  2. MetisMemory built into the model and updated with a forward pass, instead of an external vector store.
  3. FreeTokenServing Mixture-of-Experts models on personal hardware, up to the 753B GLM-5.2 on a single workstation GPU.
  4. Dream-RSIAgents that get better at exploring by replaying their own discovery history, without paying for new online runs.
  5. Object permanence in world modelsTeaching video models that things still exist when out of sight, with targeted data rather than scale.

Skills

Machine learning
PyTorch, TensorFlow, Keras, scikit-learn, XGBoost, Transformers
LLMs and retrieval
Distributed pre-training, supervised fine-tuning, embeddings, RAG, Qdrant, Neo4j
Backend
FastAPI, SQLAlchemy, PostgreSQL, Alembic, REST APIs
Cloud and infrastructure
AWS (ECS, S3, SQS, Cognito, CloudWatch), Terraform, Docker, Slurm, multi-node H100 clusters
Forecasting
ARIMA, SARIMA, LSTM, signal processing, backtesting
Languages and tools
Python, SQL, Git, Linux, Jupyter

Education

0

B.Tech, Data Science

Honours in IoT, Cloud Computing and Blockchain.

SVKM’s NMIMS, MPSTMEMumbai

2022 to 2026

Class XII, HSC

82.3%

Narayana e-Techno SchoolMumbai

2020 to 2022

Class X, ICSE

82.6%

Ryan International SchoolMumbai

2019 to 2020

About

Prithvi Singh, 22. Born and brought up in Mumbai, and still here.

Rides
A Royal Enfield Bullet 500. Mumbai to Goa, Pune, Daman and Lonavala, so far.
Games
My dad started me on old NES games, Super Mario Bros. and Contra. Now it’s anything from CS2 to GTA V, and some sim racing in Assetto Corsa.
Snooker
Whenever there’s a free table.
Discord
, usually hopping servers.
Twitch
twitch.tv/pr1tzy_, where I stream when I’m not doing anything else.
A Royal Enfield Bullet 500 parked on a red-earth ghat road under monsoon clouds
The Bullet, on a ghat road.
Prithvi from behind, sitting on a laterite ledge above a Goa beach
Goa.
Prithvi in a pink kurta on a lawn above a lake and hills
Off duty.

That’s a wrap.

For roles, projects or anything above, email is the fastest way to reach me.

prithvi30singh@gmail.com