Shivay Lamba @HowDevelop

AI Architect & Builder - Inference, LLMs, SLMs GSoC Org Admin - Jenkins | Docker Captain | CNCF Ambassador shivaylamba.me New Delhi Joined December 2018

Tweets

7K
Followers

8K
Following

1K
Likes

38K

Shivay Lamba @HowDevelop

a hour ago

@Arindam_1729 thanks for sharing Arindam

0 0 0 2 0

View Details

Continuous sensing creates memory bottlenecks on edge devices. Qualcomm & @ramin_m_h, Co-founder and CEO of @liquidai are advancing on-device AI with Liquid Context on @Snapdragon — compressing multimodal inputs for fast, low-power execution.

2 6 29 3K 1

View Details

Shivay Lamba @HowDevelop

a hour ago

@sarbjeetjohal @SantoshYadavDev @JuanPa see you there!

0 0 1 14 0

View Details

Shivay Lamba @HowDevelop

3 hours ago

@SantoshYadavDev @JuanPa I love this fun fact a lot and hope it continues 😄 Soo good to meet Juan Pa

0 0 1 39 0

View Details

Shivay Lamba @HowDevelop

3 hours ago

So cool to see @MLHacks giving out some @arduino UnoQs as part of prizes for their AI hackathon at @AllThingsOpen 🙌

0 1 20 481 1

View Details

Shivay Lamba @HowDevelop

3 hours ago

@carrycooldude morw power to u

0 0 1 21 0

View Details

Shivay Lamba @HowDevelop

4 hours ago

@SantoshYadavDev @ossia @aileenvl @JuanPa @jamesperkins @nickytonline Where am I 😁

0 0 1 84 0

View Details

Shivay Lamba @HowDevelop

4 hours ago

@nickytonline @AllThingsOpen @pomerium_io cool booth😎 - and especially cool cards

0 0 2 27 0

View Details

Shivay Lamba @HowDevelop

4 hours ago

@khushiSharma_22 @falkordb @cyberwave_com @WeMakeDevs this is so cool!

0 0 1 69 0

View Details

Shivay Lamba @HowDevelop

4 hours ago

@carrycooldude great to meet you at DevDay

1 0 1 49 0

View Details

Shivay Lamba @HowDevelop

21 hours ago

How to cut coding-agent costs by ~78.5% using Jev to route between open models All 32 selected coding tasks passed hidden tests. Plus: Beam vs Muse Spark across 20 coding tasks - comparing cost, speed, and results 👇 x.com/i/article/2110…

12 3 38 2K 17

View Details

Shivay Lamba @HowDevelop

8 hours ago

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

2 3 40 849 3

View Details

Shivay Lamba @HowDevelop

10 hours ago

@SaiyamPathak couple it with @arduino ventunoq - and on device llm inference goes brrrr

0 0 1 103 0

View Details

Shivay Lamba @HowDevelop

11 hours ago

@ilavanyajain u. r the sweetest person🫶🫶🫶

1 0 0 47 0

View Details

Shivay Lamba @HowDevelop

12 hours ago

@mfranz_on need to try the fine-tuned multi-modality we discussed with jev style models on edge

0 0 1 25 0

View Details

Shivay Lamba @HowDevelop

12 hours ago

@BhosalePratim thanks a lot Pratim, really appreciate it 😁

0 0 1 56 0

View Details

Shivay Lamba @HowDevelop

17 hours ago

Agreed, that was a real reproducibility gap, and it's now closed. The fix is Studio1-OSS/nconnect#36, released as nconnect v0.26.1. The rerun used a bundle built from fix commit 1b2d03a. v0.26.1 adds only the version bump and changelog on top, so it behaves the same, but its bundle hash differs. The two attempts are published separately, so adapter failures and model failures can be told apart: • First attempt (v0.26.0, without the fix): 15/20 passed. Four failures ended with Beam rejecting the request (“assistant turn … missing its reasoning”) immediately after escalation from DeepSeek or Kimi. Those were adapter failures: the model never got to work. The fifth was a genuine Kimi K2.6 miss. • Rerun (with the fix): 20/20 passed, with no adapter errors. Each condition's commit, bundle hash, run command and result are documented in docs/RUNNING.md.

0 0 1 105 0

View Details

Shivay Lamba @HowDevelop

17 hours ago

Glad it's useful. One honest caveat first: 78.5% is on the 32 tasks where routing paid off. Across all 50 it saved 24%, was about 3× slower, and passed 40 against 42. What worked best: well-specified tasks (validation, formatting, API changes) on the fast tier, with a strong model kept available rather than removed. Run it on a week of your own tasks with hidden checks, and compare cost per passing task, not per token. That's where you'll see whether it pays off for you.

1 0 1 26 0

View Details

Shivay Lamba @HowDevelop

17 hours ago

We checked. We counted calls that Claude Code rejected as malformed: bad arguments, an unknown tool, or a protocol error such as editing a file it hadn't read. DeepSeek V4 Flash (routed, 25 tasks): • Tool calls: 325 • Rejected: 1 (0.3%) Kimi K2.6 (routed, 10 tasks): • Tool calls: 180 • Rejected: 0 (0%) Tasks that escalated mid-task (12): • Tool calls: 366 • Rejected: 5 (1.4%) Kimi K3 (fixed, 50 tasks): • Tool calls: 765 • Rejected: 24 (3.1%) On these tasks, the smaller model didn't struggle with calling tools correctly. Where it struggled was getting stuck, and that's what the router watches for: three failed tool calls since the prompt trigger escalation to the stronger tier. 12 of the 50 tasks escalated that way, and 11 of them passed.