Rakesh Kapoor Innovation Centre · BITS Pilani, Pilani Campus

Peak engineering,on campus.

Seven technical verticals building brain–computer interfaces, quantised models on custom silicon, games that win international competitions, and the research to back all of it. Backed by Postman.

See the workJoin us
$18,500
won in prizes
16
projects on record
1
ACM publication
7
technical verticals

01 — Who we are

One of the most active technical clubs on campus, and the best resourced.

BITS X Postman Innovation Labs is backed by Postman — a BITSian-founded unicorn — and runs out of the Rakesh Kapoor Innovation Centre with round-the-clock access to hardware most undergraduate teams never touch.

What that buys is ambition. Members pick problems that take a semester rather than a weekend: an EEG headset built from the electrode up, an accelerator verified bit-exact against its own golden model, a paper through peer review. We publish our limitations next to our results, because the work is real enough to survive it.

02 — What we build

Seven verticals. One building.

One field, seven laws of motion. Nothing here is an illustration — each vertical runs the mathematics that already describes it.

01 — Signals, brains, hardware

Neurotech

Reading the body's electrical signals, and building the hardware that does it.

We build brain–computer interfaces from the electrode up — a low-cost single-channel EEG front end assembled from an instrumentation amplifier and a driven-right-leg loop, surface EMG decoders that read silently mouthed speech, wearables that turn a muscle flex into an actuator. Behind the bench sits the simulation side: a hybrid integrator that made large spiking-network models practical, and got into ACM BCB 2026. The work is published, not just demoed.

Spiral waves of colour rotating outward across the field, the way a travelling wave crosses cortex.
2
active projects
ACM
published 2026
30%
sim speedup
4
on the record

02 — Models, silicon, agents

AI/ML

Machine learning taken all the way down to the hardware it runs on.

We train models and then we make them fit. A wake-word CNN quantised from 163KB to 10KB with no measurable accuracy loss. A spike-sparsity-aware SNN accelerator written in Verilog for a Zynq-7020, verified bit-exact against a Python golden model. Research into how coding agents hand work to each other without losing the reasoning behind it. We publish our limitations alongside our results.

A smooth inward spiral at the edges that crystallises into hard concentric rings near the centre.
16×
model compression
$8.5K
hackathon prizes
RTL
to silicon
3
on the record

03 — Engines, jams, worlds

Game Development

Shipping playable games under deadline — and winning with them.

Four game jams, two engines learned from scratch, and a third-place finish in Meta's Horizon Creator Competition worth $10,000. We build physics systems, GenAI-driven NPCs, combat, and co-op VR — then organise the jams other colleges fly in for. The vertical runs on the same loop every time: learn the engine, scope hard, ship something people can actually play.

An isometric landscape of pixels being generated in front of you, with a ripple travelling across it.
$10K
Meta Horizon
6
jams shipped
3rd
worldwide
4
on the record

04 — Systems, flows, interfaces

Design

Research-led product design, not decoration.

We work the way a product team does: user research, information architecture, flows, then wireframes. The current case study is a campus sports and gym platform — QR facility access with live occupancy, slot booking, and equipment inventory — designed as two distinct interfaces for users and administrators, with the awkward cases handled rather than ignored.

Three slowly precessing rosette curves drawn over a golden-angle grid of faint dots.
3
MVP modules
2
user roles
UX
case study
1
on the record

05 — Story, edit, reveal

Video

The films that give campus events their identity.

Aftermovies, theme reveals and mascot launches for BOSM — the films that a whole campus waits for and that set the tone before an event has even started. We handle it end to end: concept, shoot, edit, sound, colour. When the video lands, the event already feels like it happened.

An expanding wavefront that develops the field from grey into colour behind it, stepping at 24 frames a second.
4
major films
BOSM
25 & 26
E2E
shoot to grade
2
on the record

06 — Users, scope, shipping

Product

Deciding what to build, and what to cut.

The vertical that asks whether the thing is worth building before anyone opens an editor: talking to the people who would use it, sizing the problem, writing the spec, and holding the scope while it is built. Most of the work is subtraction — every feature that ships is one that survived being argued out of the plan. It is the newest team here, and it is being assembled now.

A scattered field of pixels wandering at random that cools and gathers into several clusters, the weaker ones then draining into the brightest.

07 — Jams, talks, logistics

Events

Building the rooms where the rest of it happens.

We ran APOGEE Game Jam with teams from six external colleges alongside BITS Pilani — logistics, workspace, accommodation, judging, the lot. We ran API Day to get people using the tools modern software is actually built on. Events are how the Labs stay connected to everyone outside them.

A room of independently flickering pixels that gradually falls into a single unified pulse, then scatters again.
6+
colleges hosted
48h
jam format
API
Day
2
on the record

03 — Selected work

The things we actually finished.

  • NeurotechPublished · 2026

    Adaptive State-Dependent Integration for Izhikevich Neurons

    A hybrid integrator that takes tiny timesteps during a spike and long leaps through the quiet between them, making large spiking-network simulation practical.

    ~30% faster than the interpolated-spike baseline on biologically realistic sparse networks, with strict linear scaling to 10,000 neurons. Published at ACM BCB 2026.

    • NSGA-II
    • Timing wheel scheduler
    • C++
    • Spiking neural networks
    Read more ↗
  • NeurotechIn progress · 2026

    OpenEEG — Low-Cost EEG Front End

    A single-channel, battery-powered EEG acquisition chain built from parts: input protection, an instrumentation amplifier, gain staging, an anti-alias filter and driven-right-leg feedback, optoisolated from the Arduino that reads it.

    Interpretation is a frequency-domain problem, not a waveform-reading one — occipital alpha rising and falling with the eyes, and SSVEP separated by comparing FFT magnitude at the stimulus frequency against no-stimulus runs. A Pygame javelin throw takes the calibrated value as input, so a build can be checked end to end in one throw.

    • Instrumentation amp
    • Driven-right-leg
    • Arduino Uno
    • FFT
    • Pygame
    Read more ↗
  • AI/MLShipped · 2026

    Wake Word Detection on MFCC Features

    A compact CNN that spots the wake word "Marvin" in one-second clips, designed from the outset for edge deployment rather than server inference.

    0.961 precision / 0.966 recall on held-out test. INT8 quantisation cut the model from 163.2KB to 10.1KB — a 16× reduction — with accuracy effectively unchanged at 94.8%.

    • TensorFlow Lite
    • MFCC
    • Depthwise-separable CNN
    • INT8
  • AI/MLIn progress · 2026

    Spike-Sparsity-Aware LIF SNN Accelerator

    A fixed-point spiking neural network accelerator in Verilog that skips multiply-accumulates entirely for silent neurons.

    ~98% MNIST accuracy at 15–20% per-layer spike rates. The full fixed-point forward pass matches the Python golden model exactly, timestep by timestep. On-hardware timing and power measurements are still outstanding.

    • Verilog
    • Zynq-7020
    • snnTorch
    • Fixed-point
  • AI/MLShipped · 2026

    Structured Memory Handoff for Multi-Agent Coding

    Testing whether a typed decision graph beats a flat text summary when one coding agent hands an in-progress task to another.

    Across ten tasks the graph condition produced 25% fewer unsanctioned actions at identical completion rates — a real but partial win, in one of three metrics.

    • LLM agents
    • Graph modelling
    • Benchmark design
  • Game DevelopmentAwarded · 2026

    Meta Horizon Creator Competition — Survival & Resource

    A game built around a slime protagonist that consumes materials and creatures to gain abilities and overcome environmental challenges.

    3rd place worldwide in the Survival & Resource category. $10,000.

    • Meta Horizon
    • Game design
    • Systems design
  • DesignIn progress · 2026

    Sports Facilities App

    A unified platform for campus sports facilities and equipment inventory — QR entry with live occupancy, slot booking, and request tracking.

    Separate interfaces for users and administrators, with full occupancy, booking conflicts and unavailable inventory designed for rather than ignored.

    • User research
    • IA
    • User flows
    • Wireframing
  • VideoShipped · 2026

    BOSM 26 — Theme Reveal, Mascot Reveal, Aftermovie

    The full film set for BOSM 26: the theme reveal that opens the campaign, the mascot launch, and the aftermovie that closes it.

    • Cinematography
    • Edit
    • Sound design
    • Colour
  • EventsShipped · 2026

    APOGEE Game Jam

    A game jam run for teams from six external colleges alongside BITS Pilani — logistics, workspace, accommodation and judging, end to end.

    • Event ops
    • Logistics
    • Judging

04 — On the record

Judged by people who had no reason to be kind.

05 — Where we work

Rakesh Kapoor Innovation Centre.

State-of-the-art facilities, open around the clock. The point is not the hardware list — it is that no project here stalls because something was unavailable.

Photo
The Centre at night, lights on, people at the iMacs. Wide, slight motion blur, warm. Available light only.
Photo
The GPU rig with its side panel off, cables dressed, a training run on the monitor beside it.
24/7 access
The Centre does not close. Members work when the work is going well, not when a timetable says they may.
Latest iMacs
A full row of them, for everything from Verilog to colour grading.
Dedicated GPUs
Model training runs on our own hardware, not a shared queue.
Meta Quest headsets
For VR and immersive development, and for playtesting the same day you build.
PlayStation 5
Research, obviously.
Premium software
Subscriptions to the professional tools each team actually relies on.
Procurement on request
If a project needs equipment we do not have, we buy it. Nobody here is held back by a missing part.

06 — How it works

You will be out of your depth. That is the method.

An animated pixel field: a stream of loose particles at the left resolves into two knitted strands at the right, one quarter of its length for each of the four steps described below.
  1. 01

    Pick something too hard

    Projects here are scoped to a semester, not a weekend. That is deliberate — the interesting problems do not fit in a hackathon.

  2. 02

    Get taught what you're missing

    Nobody arrives knowing Verilog, or EEG signal chains, or the Horizon engine. Seniors teach it. Prior experience is not a filter here.

  3. 03

    Build it with real equipment

    Dedicated GPUs, dev boards, headsets, cameras. If a project needs something we do not have, we procure it.

  4. 04

    Ship it, and say what broke

    Competitions, conferences, campus. Every write-up we publish carries its own limitations section, because that is what makes the results worth anything.