Independent researcher and engineer. I build engines, languages, and machine-learning systems, and I work on the mathematics of what survives perturbation.
The real structure of a system is the part that survives being perturbed. One test, run across fields that are not supposed to connect. Everything in the map above is a link to something shipped, published, or provable.
One idea across unrelated fields
The research programme.
~97 preprints with companion code, spanning audio, machine learning, number theory and applied neuroscience. The full archive is a private, evolving corpus.
Robust spectral theory of dynamics
Perturbation robustness as the criterion for which structure in a system is real. Master theorems for the dominant invariant component.
The Riemann Hypothesis as observability
RH reframed as a positivity condition on a single observation operator over an admissible measurement space.
exp(x) − log(y)
The naturals, integers, rationals, reals and the primality indicator, all from one generator. Formalised in Lean 4 with no gaps left open.
Tinnitus as a phase-lagged loop
The phantom sound modelled as a predictive loop running a fraction of a beat out of step, turning on a single coefficient.
Unified Pitch-Time Field
Pitch and time as one continuous field rather than two categories, and real-time instruments built on it.
Visibility of the selector
A single inequality for when any system — a model, a journal, a market — manufactures evidence for what it was already selecting toward.
Get the research as it comes out.
I'm opening a mailing list for people who want access to the preprints and companion code as they're released. No spam, just new work when it lands.
All of this — the research, the engines, the open code — is produced independently, with no institutional affiliation. If it's useful to you, you can help keep it going.
Independent researcher. No institutional affiliation. I entered the UK as a non-citizen, where standard advice and opportunities assumed a status I did not hold, through two years under visa sponsorship marked by sustained pressure and shifting conditions.
Work produced outside academia is dismissed more often than it is checked, and authored compositions disbelieved as my own. The map above is the answer to that: every node is a real, checkable thing.
Abhishek Shivakumar · Cambridge, UK · classic homepage · built by hand
how it actually overlaps
A chronology.
Here's the honest shape of it, going back to 2014. I don't do one thing at a time. Between 2022 and 2024 I was simultaneously founding an audio company, acting as CTO of a neurotech startup, leading an AI-music engine, and running an independent research programme. That overlap is the polymathy. Click any bar for the detail.
Click a bar above to see what I was doing.
MPhil, University of CambridgeBSc, Royal Birmingham ConservatoireBased in Cambridge, UK
Everything I have built, by year
Public repos and my local, private builds together, coloured by category (each dot is a project). That OCR net in 2016 is where it started.
outside the engineering
An independent research programme.
Alongside the engineering I run a real, self-directed research effort: around one hundred preprints (102, across 13 programmes) at research.abhishek-shivakumar.com. It is not a side hobby and it is not one narrow topic. It runs from pure theory to applied audio, and it feeds straight back into what I build: the codec, the inference engine and the BCI work all came out of it.
Straight up: none of it is peer-reviewed or journal-published yet. It is an independent body of preprints and work in progress. I'd rather you know that than oversell it.
Where it belongs: audio venues (DAFx, ICASSP, ISMIR, AES), machine-learning venues (NeurIPS, ICML, ICLR) and applied-neuroscience venues. Status: 67 preprints, 10 methods, 10 engineering, 8 philosophy, plus theorem and empirical work.
Straight up: I'm a polymath. I haven't spent a decade on one narrow thing. I go deep in several at once: real-time audio, machine-learning systems, low-level C++, embedded, and the research underneath them, and I ship in all of them. If your role is hyper-specialised in one niche, there are narrower specialists than me. If you want someone who can own a problem across those areas, that's me.
Level: senior · principal · staff · lead · founding · Based in Cambridge, UK · a decade of daily C++
25 cross-platform plugins shipped to 500,000+ users; a sub-2ms neural audio inference engine; core work on the Tracktion engine beside the creator of JUCE.
Real-Time Audio / DSP Engineer
A decade of real-time C++ DSP: filters, saturation, delay, codecs, low-latency audio-over-IP with custom jitter correction.
ML Systems / Inference / Runtime Engineer (on-device, edge)
Wrote a from-scratch neural runtime that beats TFLite and ONNX by 3x at half the memory.
Senior C++ / Systems Engineer
Ten years of C++ at the metal: my own compiled language (LLVM/SPIR-V/WASM), a Zig build system, SIMD, cache-aware layout, lock-free code.
Founding Engineer (audio or AI startup)
Built an audio company to 500k+ users and 10+ engineers, and was CTO of a neurotech startup to the TechStars final round. I build the whole thing from nothing.