aakarsh
19. building around AI, mostly the parts nobody thinks about until they break.
raahAI
in progressvoice-first AI companion for real-time micro-advice during daily commutes and travel. built during the forge residency.
nazrAI
AI-powered offline marketing for india, turning weeks of manual billboard planning into faster campaign optimisation.
- →building a voice-first AI companion (raahAI)
- →thinking about how agents actually get used, not just demoed
- →curious about inference costs and token economics
- →still mostly figuring it out, same as always
output token predictor
predicts how many output tokens an LLM will generate using only the input prompt.
AI attention tracker
a zoom bot that analyses how attentive a person is during a call.
accent classifier
accent classification using MFCC feature extraction and a random forest classifier.
auto-news pipeline
n8n workflow that scrapes football news and auto-posts to twitter, tuned to read as fully human-written.
agents are easy to demo. hard to make someone use twice.
been thinking about inference as a product constraint, not just an engineering problem.
football continues to ruin my sleep schedule.
barca preseason and i'm already too invested. it's august.
why do most "AI companions" still feel like chatbots with a microphone?
building things before i fully understand what i'm building. probably the only way to learn.
i'm aakarsh, a student at bits pilani, goa. i like building software around things that are still slightly undefined: AI agents, interfaces, models, and the infrastructure underneath them. most projects start as questions rather than ideas.
outside of that: i play and watch football, i watch a lot of movies (letterboxd linked below, don't judge the ratings), i listen to music aggressively, and mostly just have fun with whatever i'm doing at the time. this whole page is kind of a journal of that. join me on the way.
if you're building something weird, i'd probably like to hear about it.