Engineering @JioHotstar, Startups, Learning, Coding, Product

Joined November 2007
Pinned Tweet
I published this blog post after a longish gap. Simple image optimisations: Optimize size Use correct dimensions Lazy load images Optimise format Use client-side cache Using CSS sprites Small images can be data images. Use fonts for icons jeet.world/common-image-opti…
1
3
25
5,296
I’m making videos about how computers and systems work, explained from first principles. If that’s your kind of thing, check out my YouTube channel and subscribe. Would love to have you along! ❤️ Link in bio 🔗
14
25
275
6,289
I am hiring Staff Backend and Web engineers for our international team (based in Bangalore). You'll work with a high-energy team and leaders to take JioHotstar's tech stack global. Sample JD: lnkd.in/gbhsKUQQ Please send your resume to [email protected]
11
28
189
17,978
I am hiring a senior software engineer for my VoD video ingestion team at @JioHotstar. This team ingests your favorite shows, movies, and Tadka microdramas. Build the next iteration of the platform with me. jiostar.wd102.myworkdayjobs.… Please send your resumes to [email protected]
13
21
2
363
40,965
I am comparing token usage, tool usage, subagents, context compaction etc. between Codex, Claude and Sarvam Code in this thread.
2
1
18
4,247
Claude used 1 tool (Write), Codex 0 (it didn't write a file) and, Sarvam used 2 (write_file and exec_command). I like the proactive testing by Sarvam/GLM 5.2.
1
377
I tried to use a prompt so that all three tools will write (but ran out of Codex limit right on cue). I will resume this experiment later :)
326
Looks like --bare doesn't work. I started normally after removing skills / mcps etc. Claude also created a file.
20
I have been using Sarvam code with GLM 5.2 for a couple of weeks now. (Haven't tried DeepSeek v4 Flash yet on Sarvam code yet). First impressions - Limits feel much more generous, GLS 5.2 is good while setting up GCP / GitHub actions /Cloudflare etc.
1
10
1,683
I have already stopped my OpenCode Go subscription temporarily because it ran out of limits really quickly. Sarvam code with GLM 5.2 is now powering all the tinkering I do.
1
4
417
Disclaimer: A bunch of my ex-colleagues are with Sarvam and I also know some VC friends who invested in them. I am also an Indian ;-) All types of biases to like Sarvam exist in my tweets :)
3
103
Building a platform at scale means running two races at once. A 100m sprint for today's challenges.
A 10,000m marathon for the future. JioHotstar CTO Akash Saxena explains why balancing speed and long-term vision is one of the hardest problems in technology and how a rocket launch became his favorite analogy for solving it. #JioHotstar #Technology #Leadership #Product #CTO #Scale #BusinessStrategy
2
2
102
Whoever thought the 'Free unlimited' Ox Alpha would help them with a good reveal later on severely underestimated the demand. (Invite-only might have been a better idea). For me, it stops responding randomly. (I have used OpenCode before, so I know that Harness is not bad).
2
8
2,812
Tried this today. Overall this is impressive (DS4 concept is mindblowing anyway). Still slower than a qwen3 coder running with vllm, but could be a worthy batch process. Running a couple of SWE benchmarks to see where we land.
⚡ If you have or want a DGX Spark, this post might be the most important one you come across this month. Serve the BEST model available (DeepSeek V4 Flash 07-31, 284B parameters) on ONE Spark @ 1,000 tok/s prefill, 59 tok/s multi-agent serving. One command to install. 🧵
1
8
1,866
Results with 256k window (not increasing right now).
1
58
I am not seeing 40+ tok/sec that others see. I am consistently between 19-30. Prefix-cache hit rate: ~93%. I need to compare the output code with Qwen3.8-27B that was running on this machine earlier to see if I like the style or raw output more from one vs other.
1
81