Yadunandan Singh
AI / Machine-Learning Engineer · India
About Yadunandan Singh
Yadunandan Singh is a self-taught AI and machine-learning developer based in India, and a B.Tech student in Artificial Intelligence & Data Science at CTITR (CT Group of Institutions), Jalandhar, which he began in 2026. He works to a "no black boxes" approach: before reaching for a library, he rebuilds the core component from scratch to understand it — a vector store with hand-written cosine similarity, an SVD recommender, a Bitcoin wallet from cryptographic primitives. His public track record is 22 merged pull requests in the c2siorg organization since June 2026, across the DataLoom data-cleaning tool and the TensorMap model builder, spanning background job execution, snapshot-based undo/redo, backend performance and CI regression benchmarking, server-side pagination, CI security, pandas transformations and React UI work.
Approach. "No black boxes": rebuilds core components from scratch (vector store, recommender, crypto wallet) before relying on libraries.
Award. Second place, CT BuildFest 2.0 cybersecurity hackathon (CT Group of Institutions, Jalandhar, 2026-10-31). Wrote the code for Veracity, a live scam-call and AI-voice detector, over 30 hours for the six-member team Coffee to Code. The story.
Education
- B.Tech in Artificial Intelligence & Data Science — CTITR (CT Group of Institutions), Jalandhar (also written as CT College, Maksudan, Jalandhar (North Campus)), began 2026 · in progress
- 18.06 Linear Algebra — MIT OpenCourseWare (self-directed) · completed
Technical skills
- Languages: Python, TypeScript / JavaScript, Rust (learning; used for Prism-RAG's vector store and bitcoin-wallet-rs), SQL
- Backend & data: FastAPI, pandas (profiling, transformations, time-series resampling), PostgreSQL, SQLAlchemy & Alembic, background job execution, REST API design, server-side pagination
- Frontend: React, Vite, Tailwind CSS, theming with semantic CSS tokens
- AI / ML: retrieval-augmented generation (RAG), vector search and cosine similarity, sentence-transformers embeddings, local LLM serving with Ollama and streaming responses, TensorFlow (CNNs), SVD-based recommender systems, linear algebra (completed MIT OCW 18.06)
- Systems & tooling: Axum, Tokio and ndarray (Rust), WebAssembly, Docker Compose multi-service setups, Git (rebase, squash, fork workflows), Bitcoin primitives (secp256k1, P2WPKH, Esplora API on signet)
- CI/CD & testing: GitHub Actions (including secure workflow design for fork PRs), pytest & pytest-benchmark (CI regression gates), Vitest, Playwright end-to-end tests
Projects
-
Prism-RAG — Privacy-first local RAG engine with a from-scratch Rust vector store, running entirely on your own machine.
Stack: Rust (Axum, Tokio, ndarray), Python (FastAPI, sentence-transformers), React, Ollama, Docker Compose
Repository: https://github.com/YADUNANDAN-SINGH/Prism-RAG
-
Veracity — Live scam-call and AI-voice detector for ordinary phone calls, running entirely offline on a laptop, with a PayPal payment gate.
Stack: Python (FastAPI, faster-whisper, PyTorch Spectra-0 / wav2vec2 XLS-R), Ollama (Llama 3.1 8B), PayPal Orders & Payments APIs, React, SQLite, scrcpy, BlackHole
Repository: https://github.com/YADUNANDAN-SINGH/Veracity
-
GlassBox — Zero-knowledge recommendation engine running client-side in the browser, with SVD compiled to WebAssembly.
Stack: Rust, WebAssembly, Leptos, SurrealDB
Repository: https://github.com/YADUNANDAN-SINGH/GlassBox-Rust-SVD-recommendation-system
-
bitcoin-wallet-rs — Bitcoin wallet built up from cryptographic primitives rather than a wallet library.
Stack: Rust, secp256k1, Esplora API
Repository: https://github.com/YADUNANDAN-SINGH/bitcoin-wallet-rs
-
Neuro-Symbolic Math Solver — Solves handwritten equations by pairing a TensorFlow CNN recogniser with a deterministic symbolic engine, so its answers are computed, not generated.
Stack: TensorFlow, FastAPI, React, Docker
Repository: https://github.com/YADUNANDAN-SINGH/Neuro-Symbolic-Solver
-
Django SVD Face Auth — Face-recognition login built on Singular Value Decomposition instead of a pretrained face-recognition library.
Stack: Django, OpenCV, NumPy, Docker
Repository: https://github.com/YADUNANDAN-SINGH/django-svd-face-auth
-
SVD-Based YouTube Recommender — Content-based video recommender using TF-IDF and Truncated SVD.
Stack: Flask, scikit-learn, SVD
Repository: https://github.com/YADUNANDAN-SINGH/YouTube-video-recommendation-model-with-SVD
-
Rust Actix Quiz Game — Quiz service with server-side score validation, built on Rust and Actix Web.
Stack: Rust, Actix Web, Vanilla JS, Docker
-
Geometric Transformation Visualizer — Interactive tool that visualises 2D matrix transformations in real time.
Stack: Python, NumPy, Matplotlib
Repository: https://github.com/YADUNANDAN-SINGH/Geometric-Transformation-Visualizer-
-
SVD Image Compressor — Image compression implemented from scratch through low-rank matrix approximation.
Stack: Python, NumPy, Flask
Repository: https://github.com/YADUNANDAN-SINGH/Image_Compression_via_Low-Rank_Matrix-_Approximation
-
AI Image Recognition Web App — CNN image classifier trained on CIFAR-10 and served through a web API.
Stack: Flask, TensorFlow, JavaScript
Repository: https://github.com/YADUNANDAN-SINGH/AI-Image-Recognition-webapp
-
Delhi NCR Property Price Predictor — Machine-learning pipeline predicting residential property prices in the Delhi NCR region.
Stack: Python, pandas, scikit-learn
Repository: https://github.com/YADUNANDAN-SINGH/Delhi-NCR-property-price-predictor
Open source
22 merged pull requests in the c2siorg organization since June 2026 (21 in DataLoom, 1 in TensorMap), all merged by maintainer @ivantha. Data from GitHub, updated 14 Nov 2026. Every pull request, with the issue it solved and the diff: https://yadunandan-singh.pages.dev/open-source/.
Writing
11 articles on the blog (RSS; full text for AI assistants in llms-full.txt).
-
Three Signals, One Verdict, and a Payment That Waits (2026-11-08)
In Part 2 of his first-hackathon series, Yadunandan Singh explains how Veracity, the scam-call detector that won second place at CT BuildFest 2.0, works. An Android phone on a real call is connected to a laptop by USB; scrcpy and BlackHole bring the caller's voice into a FastAPI backend that checks three signals at once: the caller's number (libphonenumber and a transparent rule score), the voice (an AI-voice model, now Spectra-0, scoring a 4-second window every second) and the words (Whisper, scam rules in English, Hinglish and Hindi, and a local Llama 3.1 8B that may only cite quotes found in the transcript). A weighted fusion lets any one strong signal raise the risk on its own, and every call ends in a JSON evidence report. He has since taken Veracity into the PayPal AI Hackathon, where a payment gate holds PayPal payments authorized during a risky call until the caller is verified. Lessons from the build are saved for Part 3.
-
30 Hours, One Laptop, and a Scam-Call Detector (2026-11-01)
Yadunandan Singh recounts his first hackathon, CT BuildFest 2.0 in Jalandhar, where Veracity, which he coded over 30 hours for his six-member team Coffee to Code, won second place. During a live call, Veracity checks the caller's number, how likely the voice is AI-generated (wav2vec2) and the words (Whisper plus Llama 3.1 run locally) for scam patterns in English, Hindi and Hinglish, warns the user mid-call and builds an evidence report for the 1930 helpline. Part 1 of 2: the build, the mentors' evaluation and lessons on teamwork.
-
I Used AI to Predict My University Exam Paper, So I Turned It Into an Open-Source Claude Skill (2026-10-26)
Yadunandan Singh describes how a last-minute hostel experiment with previous year question papers (PYQs) and Claude produced a mock exam that overlapped the real paper by roughly 90%. He packaged the workflow into exam-paper-predictor, a free, open-source Claude skill for university, board, entrance and professional exams in any country or language. The skill collects 5 to 20 years of past papers, maps section and topic weightage, backtests its own predictions on held-out past years before generating a mock paper, exports ready-to-use PDFs (mock paper, solutions and important topics), and scores its prediction against the real paper afterwards.
-
How to Find Your First Open-Source Project in One Evening (2026-07-24)
Yadunandan Singh explains the method he used to go from never having opened a pull request to eight merged pull requests in the c2siorg organisation (DataLoom and TensorMap) within forty days. Answer three questions on paper (your strongest language, your domain and your realistic weekly hours), turn them into one targeted AI search prompt, then spend 90 seconds vetting each suggested repository. The second half is an honest, step-by-step account of a first pull request, from local setup to review and merge.
-
ChatGPT Failed My Sister’s Live Class, So I Built a Neuro-Symbolic Math Solver (2026-05-30)
After watching ChatGPT answer a simple numerical with an essay during his sister's live physics class, Yadunandan Singh built a neuro-symbolic math solver. A React canvas sends a handwritten equation to a FastAPI backend, which normalises the image's polarity, segments each symbol with OpenCV, classifies it with a CNN trained from scratch on MNIST plus 24,000 synthetic operator images, and evaluates the expression deterministically, so the answer is exact rather than generated.
-
Stop Vibe Coding: How Elite Developers Actually Use AI (2026-04-07)
Yadunandan Singh reviews what research says about AI-assisted programming: a December 2025 study found that experienced developers keep tight control of AI agents rather than "vibe coding", and METR's randomised trial found developers who felt 20% faster were measured 19% slower. He sets out a seven-step workflow drawn from that research and Addy Osmani's process: spec first, small steps, full context, immediate tests, line-by-line review, a second model's opinion, and regular practice without AI.
-
How I Built a Face Authentication App in Django: A Complete ML Learning Journey (2025-12-22)
Yadunandan Singh describes building a Django web app that users log in to with their face, using the eigenfaces method: greyscale 100 × 100 face images are flattened, centred on the mean face and decomposed with SVD in NumPy, and logins are matched by Euclidean distance with a rejection threshold. Because one photo per user wasn't enough to train on, OpenCV generates eight synthetic variations of each sign-up photo. The post also sets out his method for learning a new framework through a project without copying AI-generated code.
-
Should You Be Scared of Agentic AI? Here’s the Reality (2025-12-05)
Yadunandan Singh explains what agentic AI is (systems that plan, act across tools and check their own results) and reports on testing Google Antigravity, launched with Gemini 3 in November 2025, by generating a portfolio site with React, Django and OTP authentication. His conclusion: agents are excellent for prototypes and internal tools but weak for long-lived production code, beginners should not use them to write their main code, and AI rewards developers with strong fundamentals rather than replacing them.
-
Don’t Be a Dumb Computer: SVD Explained With a YouTube Recommender (2025-11-17)
Yadunandan Singh explains Singular Value Decomposition by working a small example by hand: a 3 × 4 matrix of liked and disliked videos decomposes into concept strengths of 1.41, 0.92 and 0.20, and truncating to k = 2 keeps "STEM" and "entertainment" while discarding noise. He then shows how his YouTube recommender uses TF-IDF and scikit-learn's TruncatedSVD to place videos on that concept map, builds a taste profile as the average liked vector minus the average disliked one, and ranks candidates by cosine similarity.
-
I Stopped Solving Problems and Built a Tool Instead: How a Visualizer Taught Me Real Linear Algebra (2025-10-14)
Yadunandan Singh built an interactive visualiser that shows what a 2 × 2 matrix does to the unit square, and uses it to explain linear algebra through hand-worked examples. A shape is stored as a 2 × 5 matrix of vertices, so one multiplication transforms every point; a matrix's columns show where the basis vectors land; the determinant is the factor by which area scales, negative for reflections and zero when space collapses onto a line; and eigenvectors are the directions that only stretch.
-
How to Build Your Own Image Recognition Model as Your First AI Project (2025-09-24)
Yadunandan Singh documents his first hands-on AI project: a convolutional neural network trained on CIFAR-10 with TensorFlow and Keras and served through a Flask web app. The guide covers how images become 32 × 32 × 3 tensors, normalisation, data augmentation, three Conv2D and pooling blocks, dropout, and the compile and fit settings, plus a 1,563-epoch, eight-hour experiment in which strong augmentation and dropout kept validation accuracy improving instead of overfitting.
Contact and profiles
Full profile · llms.txt · about.md