AI Architect & Builder - Inference, LLMs, SLMs GSoC Org Admin - Jenkins | Docker Captain | CNCF Ambassador
New Delhi
- Tweets 7,160
- Following 1,272
- Followers 8,379
- Likes 37,662
Fun setup to work on - and support the Phonon TTS pipeline via GenieX targetting the NPU
it was fun to work on this in collab with @carrycooldude @ashwinexe @BhosalePratim
2
3
225
I got to build with @GradiumAI Phonon on launch day!
A smart mirror that sees your outfit and gives spoken style recommendations - all running locally on @arduino Ventuno Q.
@Alibaba_Qwen VLM to understand what you’re wearing
Phonon TTS to talk through suggestions
Both accelerated by the @Qualcomm_Dev Hexagon NPU
A new frontier for on-device speech.
We're open-sourcing Phonon, a suite of TTS models under 50m parameters, setting a new standard for naturalness and accuracy across languages, with the broadest inference support to date.
Phonon-31m is available under Apache 2.0.
gradium.ai/phonon🧵
3
4
26
883
So cool to see @MLHacks giving out some @arduino UnoQs as part of prizes for their AI hackathon at @AllThingsOpen 🙌
1
2
38
1,106
Glad to be back at @AllThingsOpen after 2022! I am going to present a talk on how to improve security for running agents with @Docker Sandboxes - whether it is a coding agent or a multi agent system
4
4
50
1,156
An AI coding agent becomes useful when it can investigate a problem, change the code, test the fix, and follow through to a working deployment.
That’s what I wanted to explore with the newly launched @reflection_ai Beam 🚀
Beam is built for coding, reasoning and agentic workloads. KubeTrace AI puts those capabilities to work on @kubernetesio incidents, with inference through @nebiustf.
The agent can:
→ Investigate logs, events and application code
→ Trace a failure to its cause
→ Propose code or Kubernetes manifest changes
→ Run regression tests in a sandbox
→ Build and deploy the repaired application
→ Check whether it recovered
The execution layer checks every proposed change against policy, and hidden checks decide whether the incident is resolved. You can inspect the evidence, patches, test output and failed attempts.
10
2
19
1,034
The problem with “YOLO mode” is that an agent can act with your access. A worktree won’t isolate your SSH keys, cloud credentials, or the rest of your disk.
Jozu Agent Guard puts the agent in a microVM, applies tool-call policies, limits network access, and keeps an audit trail.
I made this video with Opus 5.5 to show how the guardrails work 👇
A Practical 5 Layer Setup for Faster & Safer Agentic Coding nitter.siddz.com/i/article/210805422465…
9
6
20
1,227
Power of Local AI 😍 Running a custom Smart Mirror Demo application with @arduino Ventuno Q powered by a @Alibaba_Qwen 2.5 VL 7B model to describe what you wear, @googlegemma EmbeddingGemma 2 model for multi-modal vector search to find similar dresses, PiperTTS all targeting the @Qualcomm_Dev NPU on the Ventuno Q at IMC 2026.
Stand in front of the mirror - and get recommendations on how to improve your outfit and get recommendations on other clothing options you can try all offline.
1
4
28
2,869
Shot a quick video covering @PrimeIntellect Agent and running it with the newly launched @Alibaba_Qwen 3.8-27B model on @nebiustf using nconnect.
A few interesting features about Prime Agent:
• A persistent Python environment for executing tools and managing working context using RLMs
• A continual harness that can refine its supporting instructions based on experience
The video covers the setup, how Prime Agent works, and the coding demo with Qwen 3.8-27B.
Qwen3.8-27B is now live on Nebius Token Factory.
A compact 27B dense model for coding, research, and agent workflows, with a focus on planning and completing tasks across multiple steps.
Start building: tokenfactory.nebius.com/endp…
9
4
20
1,149
Reached highest loyalty tier of @TurkishAirlines Miles&Smiles program!
Means free business upgrades for long haul flights to North America 😁
4
33
1,388
You ask a coding agent to fix a bug. A few tool calls later, it’s installing a dependency you’ve never heard of.
Was the package name hallucinated?
Coding agents make supply-chain decisions too. Here’s how @Jozu_AI Agent Guard can help 🧵
8
3
17
999
So impressed with the Opus 5.5 explainer videos - Here's another one - Serve Models from a @Kit_Ops ModelKit on @HAMiProject
Detailed explanation of how Your model is packaged as a versioned ModelKit. KitOps pulls it from an OCI registry and unpacks it inside the Pod. HAMi schedules a controlled GPU share. SGLang loads the model locally and serves an OpenAI-compatible API. One clear path from model artifact to working inference. KitOps and HAMi.
5
4
25
898
Here's an video introduction to @tan_stack AI using a single prompt to generate a launch style video about Tanstack UI using @ClaudeDevs Opus 5.5 using @Remotion
Prompt shared below 👇
7
1
25
4,374
I built WearScout because I kept seeing great outfits and spending way too long searching for similar clothes online 👕
Give it a reference image, it uses @deepseek_ai v4.1 Flash via @nebiustf to describe the look, and Jev drives a real browser using @browser_use to find similar pieces across stores, check product pages, and return a shortlist with prices and links.
I tried it with a Tom Cruise outfit. I made a wrong call on privacy with my first demo, so I took that post down and changed the example.
Heavily inspired by and adapted from @_nancychauhans’ Hearth project 🙌
Just let Jev drive a real browser for house hunting… the numbers are insane ⚡
🏠One prompt, 4 rental sites crushed in 1m 16s for $0.0454 and it handed me 21 houses:
• 32 pages visited (all 4 sources)
• 76 browser actions, 60 clicks
• 112 model calls
5
3
16
1,788
I built a very similar 3d world like GTA Style Gameplay UI andcompared @xai @grok 4.7 against Fable 5.1, Kimi K3, and GLM 5.3 Flash and pretty surprising to see how cost efficient Grok 4.7 is. While the UI levels / 3d artifacts don't match levels of Fable 5.1, it is significantly cheaper to run.
8
18
4,590
Got hands on with @arduino Ventuno Q - time for some on device VLM model inference 🎉 with @Qualcomm_Dev Hexagon NPU inference
7
1
19
911
I got early access to the new @Cline Open Source Desktop app!
I wanted to see how it would fit into my everyday development workflow, so I tested it with the free GLM-5.3 Flash model across coding, scheduled reviews, and conversation handoffs.
For the coding task, I asked Cline to build a small Node.js task-list CLI with persistence and tests. After resolving an initial environment issue and fixing how the CLI handled corrupted JSON, it successfully passed all 15 tests.
But scheduling was the feature that stood out most.
I scheduled a background review of the code and documentation. Cline compared the README with the actual implementation and caught a genuine data-loss risk: a corrupted file could be interpreted as an empty task list and later overwritten.
That is the kind of work I don’t want to repeatedly remember to do myself.
Overall, the foundation feels promising. The interface makes the agent’s plan and execution visible, scheduled tasks can remove repetitive work from the development loop, and importing work from another coding agent opens up some interesting workflows.
Here’s my complete experience with the app 👇
Introducing Cline Desktop - a native interface for working with open weights models.
Use with ClinePass and all our free models like DeepSeek-V4.1-Flash, Musespark-1.3, or BYOK with any provider!
5
5
33
1,832
Another guide with @rudrakshkarpe also in the works using @Kit_Ops to serve models with modelkits and serve them using @HAMiProject thus providing a safe way to securely get signed models with verifiable identity
github.com/Project-HAMi/webs…
Had a lot of fun working with @rudrakshkarpe on this tutorial showing how to run @sgl_project inference on @kubernetesio using @HAMiProject GPU shares: complete with GPU memory quotas, compute throttling, and an OpenAI-compatible API @CloudNativeFdn
project-hami.io/tutorials/la…
5
3
16
1,356
Just tried the latest @deepseek_ai V4-013 model using the DeepSeek harness through @nebiustf TF Relay.
The entire setup and integration experience was seamless, and the model feels impressively quick in real-world use. It handled the task really well while maintaining a strong balance between speed, output quality, and pricing.
Two new models for coding and agent workflows are now live on Nebius Token Factory.
DeepSeek V4-Pro-0813 is the official V4 Pro release, built for coding agents that use tools, reason through complex problems, and work across multiple steps.
GLM-5.3 brings Z. ai’s latest post-training improvements for complex software engineering, from planning changes across a repository to carrying long-running agent tasks through to completion.
Try both through an OpenAI-compatible API, with Token Factory handling the serving infrastructure.
Bring your own tasks and see which fits your workflow.
Start building: tokenfactory.nebius.com/mode…
1
3
10
1,144
Spent 5+ hours debugging with GPT-6 Astra on medium, used ~40% of my usage, and still no confirmed root cause.
Building an on-device Android agent with Gemma 4 E2B + Qualcomm QMX.
live agent: ~2.6 tok/s. replaying the same request: ~11 tok/s. why the gap?😅
4
6
1,281
can agree - for me Goal Mode + Astra Medium has been effectively eating 1% usage every 5-7 minutes
1
1
183
What a wonderful start to @agentsnexus by @yashrajnayak and @sidagarwal04
Great discussions on multi agent architectures in production, security, monitoring and ofcourse on device AI agents @qualcomm_dev
1
3
37
1,035
I built an interactive 3D portfolio with GPT‑6 Astra: it turned my photo into a @Blender character, placed it inside an editorial WebGL experience, and created a revolving “Speaking Orbit” containing Earth globe mapping my talks around the world.
The most impressive thing to note about Astra is how it moved quickly between web search, computer use for Blender and code. It created the 3D assets, integrated them into the website, and repeatedly inspected and refined its output.
It also made this more refined video output.
GPT-6 Astra is now available to all Pro, Enterprise, and Business Premium users alongside the API.
I’ve been using it to build apps, LEGO models, and 3D scenes. Here are five things that changed how I use Codex 👇 nitter.siddz.com/i/article/209595222859…
7
1
25
4,007
One problem I keep running into when building AI systems: which model should I actually use for each task?
Benchmarks rarely give a complete answer. @withmartian AI Frontier looks at what matters in practice:
- quality
- reliability
- real cost
- performance across tasks
The interesting takeaway: there isn’t one model that dominates everything.
Increasingly, picking the “best model” is becoming a routing problem.
Interesting data to explore:
aifrontier.withmartian.com/
We got 46% fewer errors than the single best LLM across the 16 most used benchmarks (TerminalBench, LiveCodeBench, etc).
Here's how that's possible and what each model can achieve when used optimally (every benchmarks misses the majority of model capabilities) 👇
Interactive Site: aifrontier.withmartian.com/
Academic Paper: arxiv.org/abs/2606.26836
6
5
11
1,116
A wonderful start to the @OpenAIDevs Codex Bengaluru Meetup by @mvkaran sharing updates about Codex
@blrtechweek @yashrajnayak
3
4
22
816
Ran the same GTA-style open-world prompt through three models: Kimi K3, GLM 5.3 Flash and @ClaudeDevs Fable 5.1.
Rules: one HTML file, zero assets, everything procedural. Streets, traffic, pedestrians, day-night, enter a car and drive.
Fable 5.1 nailed it in one go. No fixes, no retries, ran exactly as generated. Also the most polished city of the three by a distance.
Kimi K3 needed a fix before it would run. GLM 5.3 Flash needed two, plus a repair pass to stop it crashing on load.
The costs are the interesting part: Kimi K3 $16.89. Fable 5.1 $15.37. GLM 5.3 Flash $5.55.
Ran @Zai_org GLM 5.3 Flash and @Kimi_Moonshot K3 on the same task: build a GTA-style procedural open-world city in one HTML file. Zero assets, everything generated from code.
GLM 5.3 Flash shipped more of the world, traffic lights, crosswalks, pedestrians on the sidewalk, seeded cities you can revisit with ?seed=N and did it for 3x less.
$5.54 (GLM 5.3) vs $16.89 (KImi K3) for the same prompt.
Both models shipped code that didn't run. Kimi's start screen couldn't be clicked, GLM had a null deref in its traffic AI. Neither caught it in 35 minutes of iterating, because the harness writes files and never opens a browser: neither model ever saw its own output.
The difference was what happened next. I handed GLM the bug report and it repaired itself in 10 minutes for $2.09, then stopped on its own. Kimi never declared itself done at all.
Ran both through @nebiustf using TF Relay.
4
1
16
5,949
Ran @Zai_org GLM 5.3 Flash and @Kimi_Moonshot K3 on the same task: build a GTA-style procedural open-world city in one HTML file. Zero assets, everything generated from code.
GLM 5.3 Flash shipped more of the world, traffic lights, crosswalks, pedestrians on the sidewalk, seeded cities you can revisit with ?seed=N and did it for 3x less.
$5.54 (GLM 5.3) vs $16.89 (KImi K3) for the same prompt.
Both models shipped code that didn't run. Kimi's start screen couldn't be clicked, GLM had a null deref in its traffic AI. Neither caught it in 35 minutes of iterating, because the harness writes files and never opens a browser: neither model ever saw its own output.
The difference was what happened next. I handed GLM the bug report and it repaired itself in 10 minutes for $2.09, then stopped on its own. Kimi never declared itself done at all.
Ran both through @nebiustf using TF Relay.
6
2
21
3,180
Friendly reminder, even if @claudeai Models are down, you can still use Claude Code with OSS models like @Kimi_Moonshot K3 using @nebiustf
In this video I show running CC with Kimi k3, with 1 step install using Nebius TF Relay.
Your favorite model with your preferred harness
6
1
15
2,877
AI coding just got a lot more accessible for developers in India with @cursor_ai Start.
They've launched a new India-only plan for ₹649/month (tax-inclusive), billed in rupees with UPI support. No more converting dollars in your head or wondering if an AI coding subscription is worth it.
The video below shows Grok 4.5 building a SaaS product from scratch. It's seriously impressive for real-world coding, and you also get access to Cursor's other models, including Composer.
If you've been waiting for an easier way to get started with Cursor, this is probably the best time.
Today we're launching Cursor Start, a new ₹649/month plan for developers in India.
Start includes generous access to Grok 4.5 and Composer, so you can plan, build, test, and ship with agents every day.
6
4
23
4,447
It's available for the fine-tuning endpoint. Will be nice to have it as a public end point as well @demian_ai
1
40
A lot of us spend time obsessing over which model to use.
I think the next optimization frontier is how that model gets executed.
If your app can ask for "GPT-5.6-level behavior" (or another reference model) and let the inference layer figure out the most efficient execution while preserving the capability and behavior your application depends on, that's a much more interesting abstraction than hardcoding a specific model everywhere.
That's why Ship caught my attention.
Instead of building and maintaining your own routing logic, the endpoint optimizes execution at inference time while targeting the same quality grade, and does it at a fixed 50% lower price. The real win isn't just a lower bill; it's reinvesting that budget into more agents, longer reasoning loops, or additional verification.
Announcing Ship: an endpoint with the highest intelligence per dollar of any frontier model.
Today, Ship makes using Opus and GPT 5.6 Sol 50% cheaper guaranteed by a quality SLA.
7
1
26
2,432
Spain won the @FIFAWorldCup
but this feels like another victory for @FCBarcelona.
From Lionel Messi to Lamine Yamal, FCB has a legacy for producing class players. And when the biggest moment arrived, it was another Barça forward, Ferran Torres scoring
2
32
919
P.S. This is my first ever official reel. 😄
I procrastinated way too long on making a reel about my experience at Google I/O this year. Part of it was that I was waiting for the new release of @reelful by @katedeyneka
Finally gave it a try, and I have to say, it's incredibly seamless to use. The way it understands the context across multiple uploaded clips and stitches them into a coherent story is genuinely impressive.
For someone like me who's too lazy to edit videos, Reelful is pretty much the perfect solution. #GoogleIO
introducing the new Reelful
Reelful is an agentic video editor in your pocket - give it an instruction and it edits for you
whenever you record your thoughts, go on events, conferences, and trips - those photos and videos never leave your camera roll. with Reelful, you can turn your media into ready-to-share clips with just a prompt!
no desktop & complex setup - it works right from your phone, where your media already lives
create your first video for free today!
3
3
28
2,097
Really loving the exclusive #BuildWeek Swag from @OpenAIDevs
Was such a great experience to host the @OpenAI Build Week hackathon in Delhi-NCR!
4
2
97
3,785
Since grok build is open source, I swapped @grok build backend for @nebiustf in one config file:
▪️ any OpenAI-compatible endpoint via [model.*] in ~/.grok/config.toml
▪️ @Kimi_Moonshot K2.7 Code as default, Qwen2.5-VL for vision
▪️ @tavilyai MCP for web search
We've open-sourced Grok Build and have reset usage limits for all users.
Open sourcing Grok Build allows anyone to support making a reliable and robust harness. Check out our code, including the Git repo for the Grok Build CLI.
x.ai/open-source
11
11
118
31,807
I am subscribed to Maintainer Brief to get weekly updates about @mesheryio
You can also get daily/weekly updates around your open source projects you maintain using the tool 🫡
Open source maintainer burnout isn't an effort problem. It's an attention problem.
The signal is buried under hundreds of issues, stale PRs, and scattered HN threads.
So I built Maintainer Brief, a weekly intelligence email that does the triage for you. powered by @nebiustf 🧵
1
1
9
671
Open source maintainer burnout isn't an effort problem. It's an attention problem.
The signal is buried under hundreds of issues, stale PRs, and scattered HN threads.
So I built Maintainer Brief, a weekly intelligence email that does the triage for you. powered by @nebiustf 🧵
8
6
35
3,573
Continuing from yesterday's @cursor_ai benchmark here's where gpt-5.6-sol-high lands on tokens per dollar?
~1,971 output tok/$ ---> perfectly in the middle. It beats gpt-5.5 and opus-4.8-thinking-high-fast, but grok-4.5 and opus-thinking-high (~2,850–2,890) edge it out, and composer-2.5-fast (~6,950) is still untouchable on value.
Performance wise, pretty similar to Fable for harder and long agent running tasks.
P.S. personal run
I have tried out grok 4.5 quite extensively on @cursor_ai in a given repository where I also tested Fable, Opus 4.8 thinking high, 4.8 thinking-high-fast, GPT 5.5 and Composer 2.5 Fast.
It was surprising to see that after multiple test runs,looking at the output tokens per dollar :
claude-opus-4-8-thinking-high: ~3,245 tok/$ (44,426 output tokens for $13.69)
grok-4.5-fast-high: ~2,955 tok/$ (15,427 output tokens for $5.22)
So Opus wins by about 10% on output efficiency.
While performance is quite similar in both.
Not quite sure, if the benchmarks are slightly different, but for my usecase both Opus 4.8 and Grok are almost similar in both performance and pricing. Grok is not really an improvement in terms of cost.
Takeaways:
composer-2.5-fast is the clear winner — most output tokens for the lowest cost (~7,182 tokens/$).
claude-opus-4-8-thinking-high is the best value among the Claude/high-reasoning tier (~3,902 tokens/$) and produced high output volume.
grok-4.5-fast-high and gpt-5.5-high sit in the middle.
claude-fable-5-thinking-high and claude-opus-4-8-thinking-high-fast are the least efficient (~605 and ~752 tokens/$) — highest cost paired with the lowest output.
P.S. this is just a personal run, and output might be different for others.
2
699
I have tried out grok 4.5 quite extensively on @cursor_ai in a given repository where I also tested Fable, Opus 4.8 thinking high, 4.8 thinking-high-fast, GPT 5.5 and Composer 2.5 Fast.
It was surprising to see that after multiple test runs,looking at the output tokens per dollar :
claude-opus-4-8-thinking-high: ~3,245 tok/$ (44,426 output tokens for $13.69)
grok-4.5-fast-high: ~2,955 tok/$ (15,427 output tokens for $5.22)
So Opus wins by about 10% on output efficiency.
While performance is quite similar in both.
Not quite sure, if the benchmarks are slightly different, but for my usecase both Opus 4.8 and Grok are almost similar in both performance and pricing. Grok is not really an improvement in terms of cost.
Takeaways:
composer-2.5-fast is the clear winner — most output tokens for the lowest cost (~7,182 tokens/$).
claude-opus-4-8-thinking-high is the best value among the Claude/high-reasoning tier (~3,902 tokens/$) and produced high output volume.
grok-4.5-fast-high and gpt-5.5-high sit in the middle.
claude-fable-5-thinking-high and claude-opus-4-8-thinking-high-fast are the least efficient (~605 and ~752 tokens/$) — highest cost paired with the lowest output.
P.S. this is just a personal run, and output might be different for others.
3
1
13
3,188
Excited to share that @krisvelkov and I will be leading a workshop on Docker Sandboxes at @WeAreDevs World Congress 2026 in Berlin! 🇩🇪
With 13,000+ developers, it's one of Europe's biggest developer conferences.
Sadly I won't be there in person, but looking forward to sharing the workshop with everyone!
8
1
37
1,298