Shreshth KapaiAI/ML/Systems

London, United Kingdom

I like ML, systems, and maths.

I'm Shreshth. Lately, I've been building AI systems, terminal runtimes, CUDA kernels, and numerical simulations.

01 / Systems and experiments

Selected work

0102 Jul 2026

Rust / GPU terminal runtime

Panea

A layered Rust terminal emulator with native platform backends, GPU rendering, shell semantics, and strict workspace boundaries.

Constraint
Terminal correctness spans VT parsing, state, PTY transport, rendering, platform integration, configuration, and security; coupling those concerns makes parity and recovery fragile.
Implementation
Structured a Rust workspace around contract crates for terminal state, parsing, semantics, transport, WGPU rendering, fonts, config, multiplexing, diagnostics, and platform adapters; enforced dependency rules with layer checks, conformance fixtures, and fuzz targets.
Result
The current alpha has separate GUI and diagnostics entrypoints, a Windows installer, local-shell smoke coverage, and an explicit cross-platform parity backlog.
  • Rust
  • WGPU
  • ConPTY / PTY
  • Fuzzing
Source
0219 May 2026

Persistent terminal runtime

wmux

A persistent cross-platform terminal multiplexer built around a server-owned session model and native PTY / ConPTY integration.

Constraint
A disconnected terminal client must not take down sessions, processes, terminal grids, or workspace state, especially on Windows without a POSIX emulation layer.
Implementation
Made the persistent server authoritative for sessions, windows, panes, layouts, jobs, commands, paste buffers, and clients; added framed IPC, native platform adapters, chunked IO, dirty-region rendering, redraw coalescing, and bounded backpressure.
Result
A Rust v1 release line targeting Windows, Linux, and macOS with explicit platform contracts, conformance and stress suites, reproducible performance baselines, and verifiable release archives.
  • Rust
  • ConPTY / PTY
  • IPC
  • Terminal rendering
Source
0321 Dec 2025

C++ numerical simulation / WebGL

Schwarzschild Geodesics

A C++ / WebAssembly and WebGL 2 visualization of null geodesics and photon trajectories around a Schwarzschild black hole.

Constraint
Interactive gravitational-lensing visualization needs physically meaningful trajectories while keeping numerical error and browser frame cost bounded.
Implementation
Implemented Hamiltonian geodesic dynamics with analytical metric derivatives, constraint stabilization, proximity-adaptive RK4 integration, ray-bundle initialization, and a WebGL 2 analysis renderer.
Result
A live browser demo with orbital camera and parameter controls plus termination, Hamiltonian-error, Doppler-shift, lensing-grid, and solid color modes.
  • C++
  • WebAssembly
  • WebGL 2
  • RK4 / Hamiltonian dynamics
Live demoSource
0421 Jul 2025

CUDA / low-latency inference

Sub-Millisecond CUDA Task Queue

A C++ / CUDA extension for latency-sensitive inference primitives and financial feature transforms, benchmarked against PyTorch implementations.

Constraint
Small GPU workloads can be dominated by launch, memory, and synchronization overhead even when the arithmetic itself is simple.
Implementation
Implemented batched GEMV, softmax, and price-vector kernels with shared-memory staging, vectorized access, manual unrolling, CUDA streams, pinned host buffers, and PyBind11 / PyTorch bindings; measured median, P95, and P99 latency with CUDA events.
Result
On the documented GTX 1650 run, the best GEMV case reports 0.008 ms median latency, 4.5x average speedup, and 7.5x maximum speedup over the reference path; results remain hardware- and workload-specific.
  • CUDA C++
  • PyTorch C++ extension
  • PyBind11
  • GPU benchmarking
Source

02 / Technical range

What I work on

ML Systems

  • AI / Machine Learning
  • TensorFlow
  • PyTorch
  • Inference runtimes
  • Evaluation and benchmarks
  • Model serving
  • Training pipelines

Systems Software

  • C++
  • Rust
  • Python
  • CUDA
  • Concurrency and IPC
  • Terminal / PTY internals

Cloud & Performance

  • Azure
  • AWS
  • GPU profiling
  • Distributed systems
  • Tail-latency analysis
  • Containers

03 / Experience and education

Experience / Education

Work experience

  1. Feb 2026 - Present

    Founding Engineer (Part-time)

    Prescient Labs · London

    • Own technical design and delivery of a production AI workflow for Colonnade Insurance, transforming real-world emails, PDFs and spreadsheets into validated structured records consumed by Dynamics 365.
    • Build evaluation and verification around LLM outputs, including schema checks, deterministic business rules, exception handling, human-review paths and representative failure cases to improve reliability and auditability.
    • Develop Python pipelines for document extraction, cleaning, normalisation, entity resolution, PostgreSQL and REST API integration, LLM processing and structured output validation across ambiguous insurance submissions.
    • Investigate model and pipeline failures such as inconsistent extraction, ambiguous mappings and latency bottlenecks; use logs, evaluation cases and measured behaviour to separate model errors from data and integration failures.
    • Provision and operate Azure infrastructure with Terraform and Docker, supporting repeatable environments, CI/CD, monitoring, secure deployment and production troubleshooting; designed changes targeting approximately 30% lower end-to-end latency.
  2. Aug 2024 - Jun 2025

    Founding Engineer

    NeuralThread AI (Acquired) · India

    • Designed and deployed AWS-based ML inference and model-serving pipelines, balancing latency, throughput, scalability and operational reliability across model architectures.
    • Built reproducible benchmarking and optimisation workflows for GPU kernels and inference components, using quantitative measurements and failure behaviour to guide engineering decisions.
    • Worked across engineering and research-style experimentation to turn model-serving ideas into reproducible implementations, architecture documentation and troubleshooting guidance.

Education

  • MSc Data Science

    University of Bristol

    Sep 2025 - Sep 2026 (expected)

  • B.Tech Computer Science Engineering

    VIT

    Sep 2020 - Sep 2024

04 / Contact

Get in touch.