Anirudhh (Arnie) Ramesh

LinkedIn - Github - Contact: anirudhh[dot]ramesh[at]gmail[dot]com

I'm a MSc (grad) CS student @ ETHZ and did my BSc (undergrad) CS at EPFL. Originally, I was born in Brussels to South Indian parents.

I'm a strong coder who can also do marketing (12.5K followers, 3M views on LinkedIn) and sales (closed CHF ~20K for my blog & hackathons). According to Naval Ravikant, that makes me 'unstoppable'

I'm set on building a research company as I love building things & I love users using it. The effort to build a small successful company vs a unicorn-scale company is similar - so might as well go big.

--

Out of necessity, I'm building up the Zurich/Swiss AI/start-up ecosystem (then I'll go back to building). You might have seen me organize AI Tinkerers Zurich or the OpenAI x Mimic x Loki Robotics Hackathon which had 300+ applicants (Google, Nvidia, YC, NASA, ...).

I'll be launching many more events/hackathons and getting the Zurich ecosystem kickstarted.

My average day

Wake-up > Code(Main) > Lunch > Code(Main) > Dinner > Study|Code(Side) > Gym|Bouldering > Sleep
*I work on what I enjoy so this is actually great

Things I've built

• GPT2 from scratch - Demo | Code

  Built for fun, by watching the Karpathy Makemore series <3. Implemented & debugged bi-gram, NN, Bengio, WaveNet, Transformer models from scratch. Implemented & trained GPT-2 on 10B FineWebEdu dataset from HuggingFace, on an 8xA100 GPU cluster from vast.ai.

• RAG for Substack - Demo | Site

  Built so my 1.5K readers can chat with my Substack newsletter. Using t3.gg stack (NextJS, Prisma, Tailwind, tRPC). Supabase Auth & DB. Pinecone VectorDB. MixedBread embeddings. 40K views, 500 prompts processed successfully.

• AI Blog Writer - Demo | Code

  Built a Python CLI which generates entire blogs (even thumbnails) from a zoom recording with a guest (& their photo + resume). I use it to semi-automate my newsletter. Uses Claude-3 Opus, contrastive Chain-of-thought, few-shot prompting. Also AssemblyAI diarizated transcription & HuggingFace BG Removal.

• PyTorch from scratch - Demo | Code | Site

  Re-built PyTorch/Micrograd from first principles. I did this during n&w while interning full-time at AWS. I followed Karpathy's Micrograd video as well as Stanford ML & DL specializations. Reviewed derivatives, backprop, I also derived sigmoid, binary cross entropy from scratch. Trained a hotdog classifier (the one from Silicon Valley xD).

Some places I worked at

• Software Engineer Intern @ Apple Cork (Localisation team)

Working with XCode, SwiftUI, evals, fine-tuned LLMs, speculative decoding for software localization/translation purposes.

• ML Researcher @ ETHZ Scalable Parallel Computing Lab

Implementing 3D Gaussian Splatting using the DaCe framework from the SPCL Lab (advised by Prof. Torsten Hoefler). Also working on large-scale distributed 3DGS, efficient gaussians streaming & client rendering.

• Software Engineer Intern @ Amazon Web Services

Worked in Network Monitoring. Lambda, S3, DynamoDB,... . Productized a data science PoC, shipped 3 services to prod handling 1M+ requests per day. Got return.

• Co-founder & CEO @ Swirl

Developed a 3D Canva competitor based on GLSL shaders, ThreeJS, NextJS. Grew 120 waitlist sign-ups. Interviewed for YC. Stopped because we started from tech (not customers).

• Software Engineer Intern @ Stacksync (YC W24)

Designed & implemented business integrations - ChargeBee (subscriptions), HubSpot (CRM) and Auth0 (auth).

• Autonomous Navigation Engineer, Marketing Lead @ EPFL Xplore (Mars Rover)

Researched and created various designs for an autonomous navigation system built with real-time SLAM & path planning with ROS1 & move_base, using sensor data from IMUs, a Kinect stereo camera and Ouster OS1-16 Lidar.

• iOS App Developer @ Sleepiz

Built their iOS app in 2 months at age 16. This helped them raise CHF 100K+ in grants.

My past projects

• Efficient, distributed sparse matrix-matrix multiplication. (5-Team project) Using MPI & C++, implemented a distributed sparse matrix-matrix multiplication algorithm. Did most design & implementation, using primitives like broadcast, Allgather(v), Alltoall(v), Allreduce. Came up with novel ideas like bloom filter for column-sharing protocol which led to 33% speedup.
• Ambitious x Driven. CS outlier/founder newsletter with 1500+ readers, ~400K views. Monetizing using job board + partners
• LLM Prompt Injection game Inspired by Gandalf, built in Python Flask. 550 users, 3.5K prompts generated.
• Arduino Vehicle Traffic Simulation (5-Team undergrad project) Built robust, short-turning radius 3D-printed cars which were used for a CV-based simple traffic simulation with a feedback control loop.

Projects backlog

• GaussianDB: World-scale Gaussian Splatting Implement an efficient 3DGS renderer in C++/Rust (using CUDA). Implement hierarchical 3DGS/Octree data structure to support continuous LoD for efficient rendering. Implement a DB structure to support multi-client, efficient streaming of world-scale gaussians. Deploy via a Kubernetes cluster.
• BitSeek: Bit-quantized reasoning model Implement BitLayer paper with 4*2b packing into int8 ops. Pre-train on GH200 cluster, then apply SFT, DPO and GRPO training to build a reasoning model. Re-write a CPU-optimized inference engine for the model using SIMD (AVX-512, ...), SSA, blocking, cache-aware, etc.
• GaussianAvatar Podcaster: Implement 3D Face Reconstruction with Dense Landmarks or GaussianAvatar paper. Likely offline-generated using SIMPLify-X (iterative-optimization fitting), perhaps mixing in FLAME + surface-attached Gaussians, dense landmarks (to fit speaker face/shoulders). Use this to generate high-quality podcasts from my Zoom guests recordings of my blog, a bit like Lex Fridman's Meta-Codec interview.
• TTS model from scratch Build a TTS similar to Sesame, Orpheus, etc. If it's good enough, use it for creating AI-generated podcasts for my newsletter. Need to dive deeper into this.

Many more things too :)

Reading backlog

• The Datacenter as a computer

Some outlier things about me

• At 14, started coding. Built a Flappy Bird clone in Unity & C#. Then self-learnt Android & iOS development, shipped apps to my high school (500+ users).
• At 15, built a school clothing enterprise. Sold 1.3K clothes, generated €20K revenue
• At 16, joined a Sleepiz, a Swiss start-up, and built their entire iOS app in 2 months. Login/Auth, view/stream heartrate, breathing, movement data from their sleep-monitoring device. They raised 6.2M in 2022.
• More things: Generated 3M views through LinkedIn. Led 2 start-up investments through s2s.vc. Founded & led the Google Developer Student Club in Lausanne.