Rahul Dev (@iamrahul_dev) header
Rahul Dev (@iamrahul_dev) profile picture

code, philosophy & space

Joined May 2021
Photo by Rahul Dev (@iamrahul_dev)Photo by Rahul Dev (@iamrahul_dev)Photo by Rahul Dev (@iamrahul_dev)Photo by Rahul Dev (@iamrahul_dev)Photo by Rahul Dev (@iamrahul_dev)Photo by Rahul Dev (@iamrahul_dev)Photo by Rahul Dev (@iamrahul_dev)Photo by Rahul Dev (@iamrahul_dev)Photo by Rahul Dev (@iamrahul_dev)Photo by Rahul Dev (@iamrahul_dev)
🧵 Building a zero-copy, lock-free Agave JSON-RPC service
2
1
13
2,209
Everyone’s rewriting everything in Rust right now. But should you? I rewrote the matching engine of my centralized exchange from TypeScript to Rust. Results: → Pure in-memory: up to 100× faster → With Redis: the gap shrinks hard… but still ~10× faster at 100k orders I measured p50/p90/p99 across book sizes and order types. Full benchmark report + code in the repo. video below 👇
12
6
1
24
792
I am looking for a job I have coded most projects from first principles and I am open to backend-focused roles reposts are appreciated and open to feedback
45
43
1
373
14,858
Coding agents burn tokens. To answer “what breaks if I change this?”, they grep and read whole files, and still miss hidden calls. At the @hydra_db hackathon I built CodeGraph It turns your Python repo into a call graph stored in HydraDB. The agent walks callers hop by hop and gets only the functions that matter, with exact file and line. ⚡ 13.8× fewer tokens than opening the files 🎯 Claude Code found 18/18 callers with CodeGraph vs 10/18 with grep 💰 $0.17 vs $0.61 for the same 12 runs, 4× faster 🔍 Learns hidden calls by watching your tests run Traverse structure. Stop burning tokens. 🧵 🔗 Live: codegraph-deploy-production.… 💻 GitHub: github.com/prabhatsin/codegr…
10
7
25
773
I built Vettly: an AI voice interviewer that reads your resume, interviews you live, and gives employers a scored report where every score links back to something the candidate actually said. Live: vettly-frontend.vercel.app/ Code: github.com/prabhatsin/Vettly How it works: → You apply and upload a resume → You join a voice call with Iris → Three agents (screener, technical, behavioral) take turns on the same call, passing context along so you never repeat yourself → After the call, each rubric criterion is scored with a quote from the transcript as evidence → Employers see the score, the quote and the replay Under the hood: Next.js frontend, FastAPI + Postgres backend, LiveKit for real-time audio, Redis for live interview state, and a worker running the voice pipeline (VAD → STT → LLM → TTS) with interruption handling. Built solo in 2 weeks. The GIF is a quick look at the UI, the system design is in the reply below, and a deeper write-up is coming next.
5
5
14
996
Here we go....
7
89
Rahul Dev retweeted
Just spoke with @micahalcorn about @JoyrideExchange. Crazy idea, ser! Thanks a lot for explaining it and answering all our questions. Wishing you the best on this journey
1
2
1
17
371
Mumbai, I am inside you
1
1
6
226
Me pretending at the airport like I read 100 books a year
1
5
195
Coming to Mumbai for Solana summit and Devcon
4
9
510
hi, i'm deepanshu. i build AI products end to end: voice agents, coding agents, and an exchange engine from scratch. looking for my next full-time role. under a minute of proof RTs mean a lot #HireDeepanshu
10
11
23
899
Rahul Dev retweeted
completed building my own @cognition infamous product aka devin ai convert your prompts directly into a product github : github.com/rshdhere/devin
13
17
1
44
1,509
Rahul Dev retweeted
See you in Mumbai, Register now - luma.com/solana-summit-india
41
38
16
879
72,456
100xdevs.com is the only official website to purchase @kirat_tw 's courses. I landed my first internship 1.5 years ago, by just following his web dev cohort. Anytime, I can vouch for his cohorts
3
9
1
91
6,013
Have been reading about sandboxing, microVMs, gVisor and firecracker since a couple of weeks, but finally booted one today. Prepared the kernel binary, rootfs and baked the rootfs image manually. Feels surreal!
2
4
157
Sandboxing... isn't just for security. Imagine this: Long running agents work for days! What if an agent wants to explore a new path? Do you wait for the first one to finish? Lot of time wasted! Answer: Snapshot the state, spin a new sandbox with the state and let it rip!
1
4
100