Skip to content
View ebinezer-rajaram's full-sized avatar

Block or report ebinezer-rajaram

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
ebinezer-rajaram/README.md

AI Inference Engineer at Perplexity in London, working on serving large language models fast and efficiently. MEng Information & Computer Engineering, University of Cambridge.

I work across ML systems and research: LLM inference, model merging and parameter-efficient fine-tuning, evaluating LLM research agents, and scientific machine learning. Currently building a from-scratch LLM inference engine.

Selected work

  • Model merging in Speech LLMs: MEng thesis. LoRA adapters in a 3B Speech LLM quietly break each other's tasks; calibrated layer-wise merging recovers single-task performance on average across 7 speech tasks without joint retraining, using 16× less training data than multi-task training. Thesis PDF
  • sciagent: a benchmark for LLM research agents in which ground truth is a structural edit to an executable programme rather than a label. It is bit-exact deterministic and content-addressed, with over 1,000 tests and mypy --strict throughout.
  • Recurrent neural operator for visco-plasticity: learns a history-dependent constitutive law from unit-cell simulations, reaching test R² 0.996 with 16.6k parameters.
  • Physics-informed networks and neural operators: PINNs for 2D elasticity and a Fourier Neural Operator for Darcy flow, with ~30× lower test error than a CNN baseline.
  • Interior-point solver: Newton, KKT, log-barrier and Phase I written from first principles in NumPy, then used to train hard- and soft-margin SVMs.
  • Cart-pole control from learned dynamics: kernel-regression dynamics and gradient-based policy search through a differentiable JAX simulator.

LinkedIn

Pinned Loading

  1. speech-llm-merging speech-llm-merging Public

    MEng thesis: calibrated merging of LoRA adapters in a 3B Speech LLM recovers single-task performance on average across 7 speech tasks, using 16× less training data than joint multi-task training.

    Python

  2. sciagent sciagent Public

    A benchmark for LLM research agents where ground truth is a structural edit to an executable program, not a label. Deterministic, preregistered, content-addressed.

    Python

  3. cartpole-rl cartpole-rl Public

    Cart-pole control from learned dynamics: periodic-kernel regression, differentiable JAX rollouts and gradient-based policy search, stress-tested under noise.

    Python

  4. recurrent-operator-viscoplasticity recurrent-operator-viscoplasticity Public

    Recurrent neural operator learning visco-plastic constitutive laws from unit-cell simulations: test R² 0.996 with 16.6k parameters, and one learned internal variable is enough.

    Python

  5. interior-point-solver interior-point-solver Public

    Interior-point solver built from scratch in NumPy (Newton, KKT, log-barrier, Phase I), used to train hard- and soft-margin SVMs.

    Python

  6. physics-informed-operator-learning physics-informed-operator-learning Public

    PINNs for 2D elasticity and neural operators for Darcy flow in PyTorch: the FNO has ~30× lower test error than a CNN baseline, and 50 measurements cut PINN error ~8×.

    Python