Building @RakshamLabs | Explorer | Engineer | ex @xAI

Joined November 2019
41 Photos and videos
Shubham Singh retweeted
Aquin CLI has 10 embedding inspection tools. - Run aquin embed-layer-drift on any text to trace how representations shift across encoder layers. - aquin embed-layer-analysis covers layer drift, isotropy, OOD separation, and consistency together. - For space geometry, aquin embed-isotropy checks how uniformly your model uses embedding directions, and aquin embed-space measures anisotropy and intrinsic dimensionality. - Run aquin embed-ood to see how well your encoder separates familiar from unfamiliar input. - aquin embed-attention shows per-head attention across all encoder layers. - aquin embed-matrix renders a cosine similarity matrix across multiple texts to sanity check clustering. For token-level work - aquin embed-attribution scores each token by how much it shapes the final embedding - aquin embed-perturbation finds which tokens the representation is most sensitive to by masking and swapping. - aquin embed-retrieval encodes a query against a corpus and reports Recall@k. Load an embedding model and all ten are ready!
Introducing Aquin CLI, the mechanistic interpretability and model analysis on tooling! Install aquin on a CUDA machine, connect a session, load a model, and run inspections, simulations, evals, or watch external training runs from the terminal. Every result mirrors live to the web dashboard at aquin.app. Get early access: aquin.app/waitlist pip install aquin
1
4
38
A lot of buzz is there why india is not able to train its own model from scratch. Look at these numbers! If you have to train a 180 Billion parameter model, and if you are deploying 2000 H100 GPUs, still it will take 229 days. And in that we haven't factored the cost and time and human resources needed to make appropriate data for training. And I want to remind that GPT 5.5 is around 9.7 Trillion (not billion). And the difference in billion and trillion is huge. These numbers are real crazy. Insane respect for @SarvamAI to train their own model. And works great for it is built. These 9.7 Trillion parameter models are not able to do engineering tasks very well like CAD and electronics. There is a gap in architecture where learning is not happening the way humans learn. We have a long way to go honestly in research worldwide and hardware and chips need to become exponentially better in less size too. Coming back to India thing, we would have to solve problem of mindset and capital first. We are too much resource constraint despite people wanting to do a lot. Are you training your own model? How and what challenges you are facing? Source of images: Claude Opus 4.8 Max Research
1
42
Shubham Singh retweeted
Last 36 hours to apply One last chance to be among the best Cohort #1 starts from July 3 thefounding.co/apply
11
18
77
10,683
Time is running up! Great opportunity for people making great things... Apply Now!!
Applications closing in 2 days ⌛️ and exactly 3 weeks for the cohort to kick off.
4
63
Shubham Singh retweeted
Jun 12
for the first time you can simulate fine-tuning before committing any compute with @AquinF03, see exactly which features strengthen or suppress, which samples hurt generalization, which layers go dead!
7
2
21
986
Experience is getting better. > Reduced the tool calls with batched changes. 5x optimisation. > Gave LLM every inch of data to peek so that it is always deterministic. > Gates and Skills for deterministic execution. > Along with Entity token translation layer which gives 10x optimisation in token and speed. MCPs can't really do all these things in themselves. Making this tool gives some superpower. Model: Claude Opus 4.8 Adaptive Prompt: (find in reply)
I have been trying to build claude code like agent tool for @adskFusion and 3D modelling, for some weeks. Here is the first preview of this tool which we call Torus internally. I will be making the code open source after some weeks. Fully MIT license so that CAD can become better usable with AI. (A lot of bugs need to be solved) This is an experiment and effort to make generalist LLM like @claudeai , Chatgpt, @grok, @deepseek_ai to be able to make deterministic and beautiful designs without fail. Just like we do coding right. Goal is simple, deterministic as well as optimised for token usage (cost matter 😅). Drop your prompts and I will burn my tokens to see how it works. (No guarantees though 😁)
2
4
426
Prompt: Build Talon, a 3D-printable unibody bottom plate for a 7-inch ArduPilot quad. Create parameters first: wheelbase = 295 mm, plate_thickness = 7 mm, arm_width = 16 mm, motor_pattern = 19 mm, stack_pitch = 30.5 mm, hole_clearance = 3.4 mm. Body: a fully constrained rounded center section about the origin, roughly 90 × 48 mm, extruded plate_thickness. One arm: a fully constrained sketch running from the center section out along the 45° diagonal, ending in a Ø32 mm motor pad whose center sits at half the wheelbase diagonal from the origin; extrude it joined to the center. Motor mount: cut a Ø10 mm wire-pass hole through the pad center, then one hole_clearance hole through the pad, rectangular-patterned 2 × 2 at motor_pattern spacing centered on the pad. Weight relief: cut one slender ellipse through the arm between pad and center. Now create a construction axis through the origin along Z and circular-pattern the entire arm — pad, holes, and relief cut together — 4 times around it. Stack mount: cut four hole_clearance holes through the center section on a stack_pitch × stack_pitch square about the origin. Strengthen: fillet the vertical edges where each arm meets the center section R8, and chamfer the plate's top outer edges 1 mm for bed lift-off. Do not scale the body — printed frames keep their hole clearances, not a shrinkage factor. Finalize, then export 3MF to /tmp/talon7.3mf.
1
58
> Heavy optimization of tool calls smart gating (especially around deletes) > Using Plan todo lists > Skill files for consistent behavior These features actually made performance better. Even with all this, a relatively simple model still struggled with proper assembly and joints. There’s still a long way to go. PS: I am using @claudeai 4.8 Adaptive for now. Have any more ideas how to make agents better in CAD and 3d modelling?
I have been trying to build claude code like agent tool for @adskFusion and 3D modelling, for some weeks. Here is the first preview of this tool which we call Torus internally. I will be making the code open source after some weeks. Fully MIT license so that CAD can become better usable with AI. (A lot of bugs need to be solved) This is an experiment and effort to make generalist LLM like @claudeai , Chatgpt, @grok, @deepseek_ai to be able to make deterministic and beautiful designs without fail. Just like we do coding right. Goal is simple, deterministic as well as optimised for token usage (cost matter 😅). Drop your prompts and I will burn my tokens to see how it works. (No guarantees though 😁)
2
76
Biggest bottleneck I solved: Fusion’s entityToken references are base64 strings. They come back in every tool response and get extremely expensive. One pattern with 81 bodies was eating ~15 KB of tokens. Each reference was costing the LLM 45–60 tokens. I built a simple translation layer that maps them to short aliases (b1, e1, s1, etc.). Result: > Token usage per reference dropped to 2–5 tokens > Execution speed improved ~5x Lot's of bug fixes and optimisation needed still. Long way to go. Drop your real world project prompts for me to test.
I have been trying to build claude code like agent tool for @adskFusion and 3D modelling, for some weeks. Here is the first preview of this tool which we call Torus internally. I will be making the code open source after some weeks. Fully MIT license so that CAD can become better usable with AI. (A lot of bugs need to be solved) This is an experiment and effort to make generalist LLM like @claudeai , Chatgpt, @grok, @deepseek_ai to be able to make deterministic and beautiful designs without fail. Just like we do coding right. Goal is simple, deterministic as well as optimised for token usage (cost matter 😅). Drop your prompts and I will burn my tokens to see how it works. (No guarantees though 😁)
1
1
87
I have been trying to build claude code like agent tool for @adskFusion and 3D modelling, for some weeks. Here is the first preview of this tool which we call Torus internally. I will be making the code open source after some weeks. Fully MIT license so that CAD can become better usable with AI. (A lot of bugs need to be solved) This is an experiment and effort to make generalist LLM like @claudeai , Chatgpt, @grok, @deepseek_ai to be able to make deterministic and beautiful designs without fail. Just like we do coding right. Goal is simple, deterministic as well as optimised for token usage (cost matter 😅). Drop your prompts and I will burn my tokens to see how it works. (No guarantees though 😁)
2
1
15
1,812
Shubham Singh retweeted
if youre into modern warfare, robotics, AI or the future of defence tech, Id love to learn from you were pivoting our startup toward autonomy infrastructure for unmanned systems and looking for honest feedback from people in the ecosys Just 15 mins lnkd.in/gArZGjUr
1
21
1,496
Agents are writing comments and structure for agents itself 🤨. Now I have to make a rule to always write according to humans. Not a great behaviour... @ai_anthropic @OpenAI @SuperGrok
2
53
India’s distribution is really weak in exports. There is no platform like Alibaba, AliExpress & Temu. Thats why we are not able to capture global demands despite having better products. Distribution is single most important thing after product. And India is losing in export story.
1
29
- World models - Photonics chip technology - Quantum computers and technology - datacenter outside earth - Robots and dark factories - … (Add more) World is changing too fast to keep up. Gotta catch up. India, a lot to catch up to…
4
26
Must try for traders out there
We’re launching Volrix. An MCP server that lets you backtest and research trading ideas directly from Claude and Chatgpt. Connect your AI agent, describe a strategy, and run backtests across index and commodity derivatives. Link in the comments, Try it for free!
1
48
Shubham Singh retweeted
We’re launching Volrix. An MCP server that lets you backtest and research trading ideas directly from Claude and Chatgpt. Connect your AI agent, describe a strategy, and run backtests across index and commodity derivatives. Link in the comments, Try it for free!
62
73
430
139,545
R&D mode... Slowly slowly, let's solve this problem of GNSS denied for good
2
75
founder first job, be shameless. sales and prompting your product ↗️ ft. @PcbCupid first esp32 with 5GHz lowest and cheapest, launching soooon! congratulations brother 🙌
6
4
45
2,165
Shubham Singh retweeted
May 31
At @AquinF03, we're continuing to make all existing evals and benchmark tools obsolete: 1/3 Custom evals: write your own scorer in Python and you get access to activations and SAE features, so you can do things like: "check whether a specific feature fired above threshold on a response" which no external eval harness can do! 2/3 Benchmark Builder now can run weight evals differently in a suite, and export results in multiple formats. 3/3 Auto-suggestions: agent observes and proactively suggests most relevant evals, with just one click to run.
2
2
6
711
That mindset is seen in every structural system in India… in even those who have money to invest and make it…
May 29
Replying to @TrustScore_1
Pessimistic mindset can also be added as a reason that kills such companies :) Neither we work to create the next Anthropic nor can we even think anyone from millions of Indians can do it either. Such pessimism and constant criticism without any solutions or work done to improve, then gets manifested as reality (be it fixing our public roads or R&D) - not in some esoteric ways, but in a material consequential way. Because it puts not just the person but everyone around in a "cannot do it" attitude. So first we need to believe "we can." Are there issues? Sure. The issues can be pointed out, spoken about, solutions offered, and if possible put our own effort to solve them to enable the next Claude or ChatGPT from India. Good news is, I know at least two companies from India that are on the verge of creating a top level model for coding and agentic tasks within a year.
59