Vrishank Aryan S

I work on a production, real-time voice AI pipeline: the prompts it runs, the audio it records and the tools the team uses to review calls.

• A prompt system that improves itself. Live call traffic is routed across 4 competing prompt variants using Thompson Sampling and Bayesian selection. An automated scorer grades every call on five rubrics: clarity, task completion, engagement, accuracy and tone.
• Audio infrastructure at scale. I reverse-engineered a third-party telephony provider's recording-access protocol and built a server-side audio proxy with HTTP Range forwarding and multi-credential fallback.
• Call review platform. A self-contained full-stack tool for reviewing calls: audio player, structured forms that auto-save, live analytics and a deterministic assignment algorithm that guarantees no two reviewers get the same call. FastAPI, vanilla JS and SQLite, packaged with Docker and deployed on Linux.

A 3D simulation in MuJoCo of a tendon-driven robotic arm: three tendons, three actuators. I'm training a reinforcement-learning environment on it, and training is going well.

Sesame is an open-source conversational robot designed by Dorian Todd. I rebuilt it on my own electronics instead of the original PCB, then wrote the firmware that ties everything together.

Real-time object detection running on-device, where compute is tight. An ESP32 turns the model's output into physical action, so the robot sees something and reacts to it.

A drone built for an ISRO-organized hackathon that simulated autonomous surveying of the Martian surface. I worked on software and electronics subsystems and on integrating them on the aerial platform under competition constraints.

A mini GPT-2 written from the ground up: custom tokenizer, token and positional embeddings, causal multi-head self-attention, residual feed-forward layers and autoregressive generation. Trained on Tiny Shakespeare.

The cl100k_base tokenizer used by GPT-4, reimplemented from scratch: subword tokenization, merge rules and vocabulary construction.

Word embeddings built without a framework, capturing semantic relationships as the base for similarity and analogy tasks.

A neural network built using only raw PyTorch tensors: no torch.nn, no autograd. Forward pass, activations, loss, backprop and weight updates all written by hand.

A spam classifier from scratch using word frequencies and Bayes’ theorem, with no ML libraries.

Perceptron, softmax regression, Gaussian discriminant analysis and SGD, written with only NumPy and the math, a repo I keep adding to.

B.Tech, Information Science Engineering, Bangalore