Backend engineer learning distributed systems & infra. now: building voice agents at scale prev: shipped iOS apps → paying users
portfolio →
Joined May 2024
- Tweets536
- Following222
- Followers128
- Likes1.2K
aryar retweeted
After years of floundering, I think I'm finally getting it: You get energy by spending it. The fuel tank metaphor is completely misleading. The body supplies energy to meet demand. The tank *expands* if you use a lot of fuel. In other words, biology is fundamentally antifragile
aryar retweeted
Enabling access to sports has long been a dream of mine.
Help us hire a tennis coach for our academy at @at_altspace
Dream job alert for Indian tennis coaches 🎾
AltSpace, a coliving and coworking community in Dharamshala, is hiring their first-ever Head Tennis Coach — to build a development programme from the ground up.
One court right now. No batches. Nothing running yet. But the vision is clear: a serious pathway for players to go state, national, and beyond — with the courts, equipment, and coaching support to back it.
What they need:
📍 3-4 years minimum coaching experience, national circuit exposure
📍 Deep technical grounding — someone who can build correct fundamentals from a complete beginner
📍 Ability to design a full programme: batch structure, progression, assessment
📍 A strong communicator who gets that coaching isn't just technical — good with parents, good with kids, good with beginners
📍 Willingness to relocate to Dharamshala
What's on offer: competitive salary, accommodation + meals covered, private session earnings, and full ownership of the programme — plus mountain views from the court that most coaching jobs in India can't match.
Not a coach? If you're a tennis lover, this is reason enough to plan a trip — stay at AltSpace, play a few sets with the Dhauladhar range as your backdrop, and see why this might be the most scenic court in the country.
Interested in the role?
Fill out the application form - airtable.com/appgVOejGe5qIcU…
OR
Write to [email protected]
good read
How we built Cloudflare's data platform and an AI agent on top of it blog.cloudflare.com/our-unif…
aryar retweeted
🌟 Announcing the BestPossible Hackathon
Teaming up with BestPossible to bring something new to the community - a hackathon around social/impact themes, building with AI
Build online, earn your way to the Himalayas. Top builders get flown in for an all-expenses-covered (stay + meals) finale.
🛠️ Build & deploy a real app, Aug 1–10, wherever you are
📤 Submit by Aug 10 · invites out Aug 13
📅 Finale: Aug 23–26, @at_altspace, Dharamshala
🏆 ₹1L Gold · many ₹10k Silver prizes · Claude pro subscriptions for quality submissions
🎫 Free to enter · can't travel? Demo online and compete for every prize
Apply → bestpossible.ai/hackathon
If it's your thing, apply - takes a few minutes. And if you know builders (alumni, dev communities, student groups) who'd vibe with this, send it their way.
hidekazu-konishi.com/entry/a…
banger read
aryar retweeted
I am excited to announce that we are officially writing a new version of Postgres. In Rust - and creating the LLVM of databases in the process.
In the span of a year, we have rewritten SQLite. Keeping the compatibility, increasing its feature set. MVCC, Types, (Live) Materialized Views, among other things. In the process of doing that, we have realized: At the end of the day, what makes SQLite special is that it compiles SQL to a database-specific bytecode. So why can't we compile *Postgres* to the same bytecode?
Turns out we can. I ran an experiment called pgmicro as a way to prove this hypothesis, and it works very well. It is time to make this official, and put the weight of Turso behind it. We shall give the world a modern take on Postgres. Wire compatible, but built on a new architecture.
We have already heard of others wanting to extend this. MySQL? Redis? the sky is the limit. What can we do if we do for databases what LLVM did for compilers? To prove how powerful the SQLite bytecode is, we are actually running DOOM compiled to the unmodified SQLite instruction set. And because Turso runs natively in the browser, you can play the game in your browser. With the database executing it.
Read the full story below! 👇
update 4/n-
voice agent - code w/o ai -
1. finished migrating the whole data layer off supabase to local postgres setup. way more control + learned a new way of writing test cases with claude.
2. reworked the prompting system - implementing logging and prompt versioning.
3. cleared the lifecycle bug. I love that i'm back to writing code manually, going through docs and fixing bugs.
just need to finish the logging set-up so i can trace everything end to end.
then we work on improve this ai agent to be more engaging and cover a lot of bases/scenarios.
code w ai -
1. made a couple of more websites - will upload them to my portfolio experiments pages... i love trying different ways to telling claude how to write reach to the inspiration...
2. working on a new project - currently call WO. will release details soon
Slowly rebuilding the foundation from first principles while running experiments with claude :)
update - 3/n -
1. Wrote the logic for custom conversations paths with the voice agent. LiveKits SDK is lit - its actually pretty intuitive to work with and to set a workflow on how to make the Agent sound more human.
2. Had set-up Tool Calling for live data query from DB about the person. This logic will get pretty interesting, when I implement match making.
3. Also revamping the backend design.
Worked on few side quests as well -
1. Tried different prompting techniques to see how far I can push claude code to replicate a website - without providing any images or website links.
Came across this post
nitter.cf/hayashimon1/status/206… and tried prompting claude to make it - came pretty close :)
Another side quest -
I've always had some difficulty trying to read and pronounce Sanskrit Words.
I already had a personal website where I was slowly curating info about the mantras, shlokas and the steps to follow.
Then I came across this post and ran this on Google Colab to generate MP3s for Vishnu Sahasranama as a test.
Its sooo good.
Huge respect to @prathoshap for building this and for making the weights open.
A 15-year-old dream has come true today. I started a PhD with the dream of creating a system that chants any Sanskrit shloka perfectly.
And here I am opening sourcing 𝐕𝐚𝐠𝐝𝐡𝐞𝐧𝐮 - 𝐀 𝐯ṛ𝐭𝐭𝐚 (𝐦𝐞𝐭𝐞𝐫) 𝐚𝐰𝐚𝐫𝐞 ś𝐥𝐨𝐤𝐚-𝐭𝐨-𝐜𝐡𝐚𝐧𝐭 𝐭𝐞𝐱𝐭-𝐭𝐨-𝐬𝐩𝐞𝐞𝐜𝐡 (TTS) 𝐬𝐲𝐬𝐭𝐞𝐦 𝐟𝐨𝐫 𝐒𝐚𝐧𝐬𝐤𝐫𝐢𝐭. This is the world's first vrutta-aware, open-source TTS for Sanskrit Chanting.
update - 3/n -
1. Wrote the logic for custom conversations paths with the voice agent. LiveKits SDK is lit - its actually pretty intuitive to work with and to set a workflow on how to make the Agent sound more human.
2. Had set-up Tool Calling for live data query from DB about the person. This logic will get pretty interesting, when I implement match making.
3. Also revamping the backend design.
Worked on few side quests as well -
1. Tried different prompting techniques to see how far I can push claude code to replicate a website - without providing any images or website links.
Came across this post
nitter.cf/hayashimon1/status/206… and tried prompting claude to make it - came pretty close :)
Was able to finish this flow -
1. Data going from - messages -> consolidate transcripts -> extract traits -> find user goals, likes, dislikes, contradictions -> save in db.
2. This is injected into next users conversation.
Makes the interactions interesting and genuinely user oriented.
Next steps -
1. Think of how to use embeddings for this scenario.
2. Implement prompt versioning + rubric based evals.
3. Implement redis and queue to real time system and run test on couple of users.
4. Cover all scenarios of user's interaction with AI and write code for it using LiveKit SDK.
Was able to finish this flow -
1. Data going from - messages -> consolidate transcripts -> extract traits -> find user goals, likes, dislikes, contradictions -> save in db.
2. This is injected into next users conversation.
Makes the interactions interesting and genuinely user oriented.
Next steps -
1. Think of how to use embeddings for this scenario.
2. Implement prompt versioning + rubric based evals.
3. Implement redis and queue to real time system and run test on couple of users.
4. Cover all scenarios of user's interaction with AI and write code for it using LiveKit SDK.
UGC IS 8%
11 of 19 video ads are creator-produced. But that's 8% of the total account.
Obvi built 70% of its account around video storytelling.
FCT uses UGC as a thin supplement layer on top of algorithm-driven DCO.
One standout:
a reverse-psychology hook -
"DO NOT BUY THIS T-SHIRT!!!" - running in 2 placements.