Shaurya Raghuvanshi

01 Work

Three things I actually shipped. Each one started from an idea I wanted to exist, not from a tool I happened to have.

NeuroLab

React 18 · TypeScript · Vite · SVG

A microscope for neural networks. Four architectures you can drag, edit, and watch think — with no ML library underneath any of them.
Every forward pass, every backprop step, every spike is computed from scratch. The engine directory is the whole point: hand-written activations, Xavier initialisation, MSE deltas, and four separate simulators.
- Dense MLP — full backpropagation and gradient descent, trainable in steps of 1 / 10 / 100 / 1000.
- Elman RNN — hidden state carried across time steps, recurrent weights drawn as dashed self-loops.
- Spiking (LIF) — membrane potential animates inside each neuron; spikes flash, refractory neurons go dark.
- 1D CNN — the convolution kernel slides across the input in real time over live feature maps.
Neurons drag anywhere and the bezier connections follow. Click a neuron to edit its bias, click an edge to edit its weight, export the whole network as JSON and pick it back up later.
Repository ↗
RankFile

Vanilla JS · PWA · No build step

A chess square-recognition speed drill. No framework, no dependencies, no bundler — it installs to the home screen and runs with the network off.
Built for board vision: name the square, find the square, or call its colour, against a clock or three lives.
- Weighted randomisation — after the first round it biases prompts toward the squares you miss or find slowly.
- Reaction-time heatmap — the board recolours green to red so you can see exactly where your blind spots sit.
- Three difficulties — coordinates shown, coordinates hidden, then hidden and the perspective flips at random.
- Tactile by design — every control presses in, haptics tick on mistakes, and all sound is synthesised through the Web Audio API rather than shipped as assets.
Lifetime stats, per-mode personal bests, and full keyboard play all persist locally. A service worker caches the app shell, so the whole thing works offline.
Repository ↗
labrag.pro

Python · Qdrant · MCP · Cloudflare Tunnel

A private retrieval system for my own notes, self-hosted on a Mac mini and reachable from inside a conversation.
Notes, project references, and an Obsidian vault are chunked, embedded, and served from hardware I own. It is exposed as an MCP endpoint through a Cloudflare Tunnel, so a model can query the corpus mid-answer instead of me pasting context in by hand.
- Semantic search returns ranked chunks with a source path, a title > header breadcrumb, and a cosine score.
- Live data sits on the internal SSD; the external drive takes the backup.
- Runs always-on alongside the rest of the fleet — see The Lab.
71 documents 301 indexed chunks 1 box, mine
Private endpoint

02 Research

A long-running project on brain-inspired efficient AI. The premise is simple: intelligence at current energy cost does not scale, and the fix is unlikely to be a bigger matrix multiply.

Efficiency gains get spent, not saved — so the interesting question is what a fundamentally cheaper substrate for computation looks like.

Neuromorphic

Leaky integrate-and-fire neurons, spiking networks, event-driven computation. The strongest pillar, and where the depth goes.

Approximate

Quantisation, pruning, approximate arithmetic — buying large efficiency wins against small, bounded accuracy loss.

Photonic

Optical matrix multiplication and the noise characteristics that decide whether it is ever practical.

Temporal memory

RNNs, LSTMs, and Legendre Memory Units. The cross-cutting foundation the other three keep leaning on.

Six phases

INeural network foundationsMLP and CNN on MNIST, written out until they need no reference.
IITemporal memoryRNN, LSTM and LMU delay tasks; LTI systems and memory representation.
IIINeuromorphicLIF simulation, spiking networks on MNIST and N-MNIST, event-based data.
IVApproximate computingQuantised MLP, pruning experiment, approximate multiplier demo.
VPhotonic computingOptical matrix multiplication and optical noise, in simulation.
VIFinal integrationA spiking chatbot — intent classifier, then generative, then SpikeGPT inference.

Implementation runs on snnTorch and Norse, on the RTX 5070 box, with the reference corpus served off the Mac mini.

03 The Lab

Not general self-hosting — a compute stack built for autonomous AI work and retrieval. Everything is wired, everything is headless, and the whole fleet is driven over SSH from a laptop.

Desktop

Heavy compute, idle by default

i7-12700KF · RTX 5070 · 32 GB DDR5

1 TB Gen4 NVMe dual-booting Ubuntu and Windows, untouched. Proxmox installs to a separate 250 GB SATA SSD.

Mac mini M1

Always-on

16 GB unified · 1 TB external

Serves the RAG stores and does the constant data shuffling. Also the backup target for the desktop.

XPS 9310

The terminal I sit at

i7-1185G7 · 16 GB · Linux Mint

No monitor on the desktop, none on the mini. Everything happens over SSH from here.

Virtualisation

Proxmox on pve.labrag.pro, hosting three Ubuntu Server 24.04 nodes cloned from a single template and bridged onto the physical LAN via vmbr0. Three is the smallest number where scheduling and rescheduling stop being hypothetical.

k3s — control plane on node one, workers on two and three.
CI/CD — Gitea and a runner, building the RAG service and deploying it on push.
Ansible — three identical hosts is the point where a playbook finally beats manual SSH.
Failure testing — snapshot, break it, restore. The one thing the always-on mini can never be used for.

Running your own compute stack is ownership, not nostalgia. Infra skill is not irrelevant to AI work — it is most of what makes AI work run.

04 Teaching

I run Young Engineers Halifax — an international STEM enrichment franchise teaching engineering to ages four to fifteen through hands-on LEGO® models. My role grew out of instruction and into operations: scheduling, staff, systems, parent communication, and the numbers.

A Young Engineers session in progress — A build session, mid-experiment

Engineering education should build thinkers, not button-pushers. The win isn’t a finished model; it’s a child who reasoned about why it works and what to change.

I took it on for the experience and the cause rather than equity, and because the operation around the teaching is a genuine systems problem — run on effort where it should be run on process. Four pillars hold it up.

Operations

Make the location run without me holding it together.

Growth

Fill more seats, and keep them full.

Curriculum

Protect the quality that earns the word of mouth.

Ops software

Build custom tools only where the existing systems fall short.

The spiral

Each tier reuses and extends the one before it, so a child has somewhere to go next — which is also, not coincidentally, the retention model.

Program	Ages	Focus
AlgoBuddy	4–6	Early coding and logic foundations
Big Builders	4–6	Building, planning, motor skills
Bricks Challenge	6–10	Forces, energy, torque, leverage
AlgoPlay	6–10	Playful algorithmic thinking
Galileo Technic	7–10	Transmissions, centre of gravity, ratchets
RoboToys	9–12	Motorised and robotic builds
Robotics	12–15	Programmable robotics
AlgoC	13–18	Text-based coding in C

05 About

I knew at eight that I loved computers. I have always counted finding that direction early as luck more than merit.

I study computer engineering in Halifax. I work top-down: start from the idea and bend the tools toward it, rather than letting whatever tool is at hand decide what gets built. I learn from structured references rather than scattered tutorials — usually by typing a reference implementation out and understanding it as it goes down.

A lot of what I do sits at the seam between two things that are normally kept apart: research and the machine it runs on, algorithms and the operation around them, engineering and the reason for doing it. I like it there.

At the desk

Elsewhere

Chess: About 1700 on chess.com, almost entirely blitz, across roughly 5,700 games — and no opening theory whatsoever. Volume taught me to avoid bad positions; it is pattern recognition rather than calculation. Trying for 2000 the same way, though the clock is starting to argue.
Aviation: A long interest in the engineering: the SR-71, Concorde, and the energy-density wall that keeps electric aircraft small. I fly a DJI Mini 3 in the meantime.
Elsewhere still: Indian spiritual geography and cosmology — Kailash, Kashi, the Shaivite traditions — plus security and cryptography fundamentals. Different rooms of the same house.