hey Paras! 👋
i've applied for the Lossfunk residency and wanted to reach out directly. i'm currently a founding engineer at Outspeed (realtime voice AI infra startup, check outspeed.com). [resume]
why i'm excited about Lossfunk?
i'm deeply curious, especially about model serving and inference, and i love understanding how things work internally. [here's the proof]. i also wanna be among people where i'm the dumbest one.
my 'from scratch' projects (AI related)
-
model serving from scratch
a toy model server built from scratch using FastAPI. has request queueing, batched inference, independent sequence completion, and token streaming via SSE. also supports batched inference for whisper-large-v3-turbo. built this to understand how systems like vLLM work internally (i'm not even close). next steps: dynamic batching and KV caching. this is what i wanna spend my time on at Lossfunk
[source code] -
numpy from scratch in Go
i was implementing Neural Networks from scratch in Go and as a byproduct of it, i created my own version of NumPy for tensor operations like +, -, *, /, matrix multiplication, transpose etc. it also supports broadcasting.
[source code] -
tensor visualizer
a tool that helps you visualize multi-dimensional arrays. uses HTML canvas for rendering the image. i wrote this when i was trying to understand neural networks and tensors in pytorch. it's the most creative and original project i've ever built.
[source code] | [live demo] -
ai answer engine
built a quick & fun answer engine inspired by perplexity & scira because why not. used qwen3-32b hosted on groq & used google's custom search JSON api for search results. it shows sources and model's reasoning in the UI, while citing the sources in the answer.
[source code]