Harsh Agarwal

Software Developer

[Amit Vadi @vadiamit]

incredible week in India.

two things stood out:

1) the moment for builders is now. I’ve spent this year with builder communities around the world and the same energy keeps showing up: people building for their own local use cases, on their own terms. India might be the strongest example yet.

2) we’re still so early. many innings left in this game

[Harsh Agarwal @harsh_agw]

@vadiamit Had an amazing time. Thanks for making it happen! 🙌

[Harsh Agarwal @harsh_agw]

@jintuisbusy Thanks!

[Harsh Agarwal @harsh_agw]

Still can't believe it. We won 🥇 1st place at the @GoogleDeepMind × @cerebral_valley Hackathon Bangalore! 🎉

We built an AI-powered RPG where worlds, assets, dialogues, and quests are generated in real time using @NanoBanana. Every playthrough is unique.

[Harsh Agarwal @harsh_agw]

built < 6h. Here's a breakdown of how everything came together behind the scenes 👇

[Parth @mittalparth_]

lots of ppl asked about how we did it, so here we go!

the why:
> the idea was to push the limits of the new @NanoBanana model which is fast and cheap. building a game came to mind as they are a good representation of pushing creative limits.

the how:
> you type an idea
> gemini-3.5-flash builds a game bible - clues, npcs, the story, rooms, street, etc. the bigger game mechanics like no of clues & npcs are always fixed.
> nano banana paints a top-view street image (the background you walk on)
> nano banana runs again and draws magenta outlines around every object (buildings, paths, trees, etc.). this is done so the next model can “see” where things are
> gemini-3.5-flash looks at the clean street + the outlined copy and marks hotspots + edges (where you can walk / enter / talk)
> at the same time, nano banana draws your player character
> also in parallel: the 3 story rooms (interiors) start building in the background while you walk, so opening a door feels instant. some buildings on the street are just flavor and don’t open into a real room
> the world is infinite - meaning if you enter the edge of a screen, it will create another screen with the previous frame as a reference
> we now have all the elements to play. all of this is rendered in an HTML canvas in the browser, where you play.

the gameplay:
> every run is one mystery with one goal
> enter a room, talk to an npc - each one guards 1 clue
> gemini-3.5-flash writes the dialogue, then gemini tts speaks it out loud
> explore, move around, pick up items, notice local nuances, engage with the npcs until you’ve got all 3 clues!

[Harsh Agarwal @harsh_agw]

@ahmedfahim21_ @GoogleDeepMind @cerebral_valley @NanoBanana Safe to say I'm hooked now 😂

[Harsh Agarwal @harsh_agw]

@ahmedfahim21_ @GoogleDeepMind @cerebral_valley @NanoBanana @GeminiApp still can't believe it we pulled it off 🏆🏆

[Harsh Agarwal @harsh_agw]

@sampathbalivada @GoogleDeepMind Congrats🔥🙌

[Harsh Agarwal @harsh_agw]

@mittalparth_ @GoogleDeepMind @cerebral_valley @NanoBanana We cooked 🔥

[Parth @mittalparth_]

update: we are in the finals!!!!

[Harsh Agarwal @harsh_agw]

@TheGracia_here This is awesome. I went down the same rabbit hole last year and built my own VM in Go.

https://t.co/qJMQa67azx

[Harsh Agarwal @harsh_agw]

Introducing Viri

Checkout the code - https://t.co/Cr0D711UEr
Playground - https://t.co/b916UEtK7G

Viri feature set 🧠⚙️
– functions with closures
– classes with methods and inheritance
– lexical scoping and block scopes
– loops, conditionals, and expressions
– modules with import / export and resolution
– runs as both a tree-walking interpreter and a compiler + bytecode VM (~3× faster)

Follow along. Actively building and iterating.

[Harsh Agarwal @harsh_agw]

@VadimStrizheus building postings - search engine from scratch.
https://t.co/fdAm1aFvz8

[Harsh Agarwal @harsh_agw]

@buildinpublic Working on building my own search engine
https://t.co/dYJfvDJ9cq

[Harsh Agarwal @harsh_agw]

@trikcode Can we even call it vibe coding if they have real programming skills?

[Harsh Agarwal @harsh_agw]

Working on a search engine to really understand how it works 🧠

One optimization that actually moved the needle:

Earlier, boolean queries built full in-memory lists for each term upfront. That meant decoding doc IDs plus extra data and allocating memory before knowing which documents would even survive AND / OR.

Changed the execution model:

Boolean queries now operate only on document IDs using bitmap-based set operations. All the heavier work (scores, lookups, extra decoding) happens only after the final document set is known.

Net effect: less decoding, less memory pressure, simpler execution flow and faster query execution

Building in public. Feedback and suggestions welcome 👇

[Harsh Agarwal @harsh_agw]

Building a simple search engine called postings.🔍

Just published my first article on how it works.
https://t.co/ByrE6DOkbI

[Harsh Agarwal @harsh_agw]

https://t.co/C6O2hV18pt

[Harsh Agarwal @harsh_agw]

https://t.co/C6O2hV18pt

[Harsh Agarwal @harsh_agw]

Building a search engine from scratch in Go.

Repository: https://t.co/fdAm1aFvz8

A detailed post on the internal workings is coming soon.

[Harsh Agarwal @harsh_agw]

Added fuzzy search, regex search, and prefix search⚡️

Check it out - https://t.co/zczayQwoJ6

All of them fall out naturally from the FST-based term dictionary. Fuzzy queries use a Levenshtein distance automaton, regex queries use a regex automaton, and prefix queries are just a range scan over the FST to collect matching terms.

Once terms are collected, the rest is straightforward: fetch postings and aggregate docs. The heavy lifting is done by the FST.

[Harsh Agarwal @harsh_agw]

Added fuzzy search, regex search, and prefix search⚡️

Check it out - https://t.co/zczayQwoJ6

All of them fall out naturally from the FST-based term dictionary. Fuzzy queries use a Levenshtein distance automaton, regex queries use a regex automaton, and prefix queries are just a range scan over the FST to collect matching terms.

Once terms are collected, the rest is straightforward: fetch postings and aggregate docs. The heavy lifting is done by the FST.

[Harsh Agarwal @harsh_agw]

Introducing postings 🔎

A search engine built to reason about modern full-text search architecture.

https://t.co/rzwnk1PwyF

The system follows an LSM-style design: writes go to a mutable in-memory layer and flush into immutable on-disk segments. Segments are memory-mapped (mmap) for efficient reads, FST-backed dictionaries drive term lookups in the inverted index, and per-segment bitmap tombstones hide obsolete documents.

In parallel, reading Introduction to Information Retrieval and studying @blevesearch’s production search engine design.