Sankalp Singh
I build simulations and then test whether their results hold up. Most of my work sits between machine learning and mathematical modelling: tracking a drone without GPS, simulating robots on Mars, and asking when a neural network’s answer can be trusted. I have written two research preprints, and one of them reports a negative result. I am applying for an M.Sc. in computational science or computer science, starting 2027.
Education
B.Tech, Artificial Intelligence and Data Science
Guru Gobind Singh Indraprastha University, Delhi
Sep 2022 to May 2026. First two years in core mathematics and physics. Dissertation: CompliSense, supervised by Dr. Ashish Joshi.
Research (preprints)
A Lead Time Is Not a Detection
Preprint. S. Singh, 2026.
Surrogate Trust Audit
Preprint. S. Singh, 2026. Code and full audit trail are public.
Experience
Backend Developer / ML
Product Manager Accelerator, USA, remote
- Built the StylePilot AI fashion platform: a PostgreSQL schema with 8+ models in SQLAlchemy, and a FastAPI backend.
- Added pgvector similarity search and a computer vision pipeline using Gemini, SAM and CLIP.
- Shipped a recommender and a RAG chatbot on top of it.
- The team placed 2nd at the showcase.
AI/ML Intern
Prodigal AI, Delhi
- Dockerised a startup health-scoring app built in Streamlit.
- Built RAG and OCR pipelines.
- Wrote Apache Airflow workflows, in a team of four.
Student Research Intern, AI for Drug Discovery
GGSIPU, Delhi
- Led the machine learning work in a team of five.
- Built RDKit molecular fingerprints for candidate compounds.
- Trained Random Forest and ANN models to predict bioactivity for colon-cancer compounds.
Selected projects
SkyScout
A Mars scout drone that maps hazards from the air and hands safe routes to a rover.
- 218 CAD solids across both vehicles
- 35 / 34 Rover links and joints
- 21 / 20 Drone links and joints
build123d / OpenCASCADE / URDF / SRDF / SDF / Genesis / MuJoCo / NumPy / Blender 4.5
Adversarial Swarm Defence Sim
A GAN learns to disguise a drone attack. A detector learns to catch it anyway.
- 6.0 Generator output scale, after a bare Tanh reached under 40% of the real feature range
- 55 vs 145 Epoch where mode collapse arrived with minibatch-std and DiffAugment, against without them
- 560 Samples in the dataset, 10 features across 99 timesteps
WGAN-GP / TCN / Multi-sensor fusion (radar, optical, acoustic, RF) / Three.js / Python / WebSocket
inkless
Markdown to a typeset PDF, with the PDF format written out byte by byte. No dependencies.
- 0 Third-party dependencies. Python standard library only
- 151 Test functions, 52 of them table-driven with subTest
- SHA-256 A rebuild in an empty virtualenv is byte-identical to the committed example
Python standard library only / Hand-written Markdown parser / PDF object model and writer / PNG chunk parser / zlib / unittest
CompliSense
An AI legal-compliance platform for Indian small and medium businesses.
- 84.8% BERT NER macro F1
- PostgreSQL 16 + pgvector Document store and similarity search
- Neo4j 5 Regulation graph
BERT / XGBoost / SHAP / PostgreSQL 16 / pgvector / Neo4j 5 / Supabase / Render / Vercel / AuraDB
The other 7 are on the missions page: Inertial Ghost, Supernova Simulation, Global Weather Forecasting, HireScope, regexray, GLCM Texture Analysis, AI for Drug Discovery.
Skills
- Languages
- Python, SQL, JavaScript, Java, GLSL
- ML and deep learning
- PyTorch, TensorFlow, scikit-learn, XGBoost, LightGBM, SHAP, GANs (WGAN-GP), TCNs, Transformers (BERT), FNO, PINN
- Scientific computing
- NumPy, SciPy, Kalman filtering (ESKF), Factor graphs, IMU pre-integration, Spectral methods, Crank-Nicolson, A* planning
- Robotics and simulation
- URDF / SRDF / SDF, build123d CAD, Genesis, MuJoCo, Godot 4, Three.js, Blender (bpy), OpenCV, GLCM, Bayesian log-odds mapping
- LLM and data
- RAG, FAISS, pgvector, sentence-transformers, spaCy, OCR (Tesseract)
- Backend and infrastructure
- FastAPI, Flask, React, PostgreSQL, Neo4j, Redis, SQLAlchemy, Docker, Apache Airflow, Azure, Git, pytest
- Research practice
- Pre-registration, Null models, Multi-seed studies, Convergence tests, Filter consistency (NEES)
Every one of these links to a project on the skills page.
Languages and certifications
- Hindi, Native
- English, IELTS 8.0
- German, Learning, A2 target
- Aeromodelling Certification