Sign in

Deb Sahu

@debashishsahu.com
170 followers 234 following 274 posts

🧲NMR/Biophysics PhD🧬Science🧪Nerd🫙Freshnox - Cofounder & CTO🧊S3 Tech Innovations - Founder & CEO🍛Foodie📡IoT enthu pattani👨🏽‍💻Coder👨🏽‍🔧Tinkerer/maker▶️Ex-YouTuber〽️Amateur Trader⚡️EV🚗Dog Dad🐶Smart Home🏠 🌏 social.debashishsahu.com 🗣️ Opinions are my own 🫢

PostsRepliesMedia
Deb Sahu @debashishsahu.com · 08/08/2026
This technology has implications in AR/VR world, as well as in historical preservation of monuments and sites by creating their digital twin. Play around this on your phone or browser: splats.earthbytelab.com/hickorywood/
splats.earthbytelab.com
EarthByte Lab
000
Deb Sahu @debashishsahu.com · 08/08/2026
Is this a drone footage? No! It a #DigitalTwin of Hickorywood park, a 3D world reconstructed by using #GaussianSplats using some fancy #AI technologies developed in collaboration with @earthbytelab.bsky.social It's now a world with colliders that one can walk through like a video game.
100
Deb Sahu @debashishsahu.com · 26/07/2026
The most useful AI thing (if you have a Trader Joe’s nearby) I have seen in a while: github.com/SGShuman/tjs...
github.com
GitHub - SGShuman/tjs-meal-planner
Contribute to SGShuman/tjs-meal-planner development by creating an account on GitHub.
000
Deb Sahu @debashishsahu.com · 16/07/2026
First they rewrote SQLite in rust, now Postgres is in the crosshairs turso.tech/blog/a-new-m...
turso.tech
We're building Postgres in Rust. Using the LLVM of databases
Turso is becoming the LLVM of databases: one modern, reliable Rust core with many SQL frontends. After rewriting SQLite, we are now building Postgres.
000
Deb Sahu @debashishsahu.com · 07/07/2026
So glad to be able to assist with NMR portion of this paper from Arul’s lab www.pnas.org/doi/10.1073/... #nmrchat
pnas.org
A ligandable PNT domain establishes ERG as a directly targetable oncogenic driver in prostate cancer | PNAS
The TMPRSS2:ERG gene fusion, present in approximately 50% of prostate cancers in patients of European ancestry, drives oncogenesis through aberrant...
021
Deb Sahu @debashishsahu.com · 29/06/2026
One of the reasons I cancelled my Oura membership is the cloud dependency, someone deciphered the BLE protocol: github.com/Th0rgal/open... My past weekend project: based on this protocol, I was able to build a macOS app that does data ingestion locally from my Oura ring... on to some ML training!
openoura screenshot
000
Deb Sahu @debashishsahu.com · 08/06/2026
Apple is also excited about the same stuff as I am! Let’s see what’s in store at WWDC’26. radiancefields.com/apple-maps-f...
radiancefields.com
Apple Maps Flyover Is Getting a Gaussian Splatting Upgrade - Radiance Fields
Apple Maps will be receiving a gaussian splatting makeover later this fall.
010
Reposted by Deb Sahu
Magnetic Resonance News @nmr900.bsky.social · 05/06/2026
Anthropic (June 5, 2026) Claude vs. ChemDraw and MestReNova on NMR prediction and structure elucidation www.anthropic.com/research/mak... #NMRchat 🧲
anthropic.com
Making Claude a chemist
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
052
Deb Sahu @debashishsahu.com · 30/05/2026
Two weeks of math optimization & 10min 360 camera drone footage orbiting my house: an amazing “Gaussian Splat” of my house! Unfortunately I won’t be sharing the ply file to play with for privacy reasons, but definite progress. Now will add LiDAR data, remove sky from input & look for improvement.
000
Deb Sahu @debashishsahu.com · 18/05/2026
TLDR - Python is great for prototyping the math for validation. Move to go/c/c++/rust to speed up real-time on-the-fly calculations. All these can be encapsulated as binaries that can use Python as a wrapper.
000
Deb Sahu @debashishsahu.com · 18/05/2026
I read a bunch of papers on a math calculation for one of my sensors, loved to combine all of them. Claude systematically combined in python (in 5 days) after breaking down the problems into smaller chunks. Then we spent 8 hours rewriting the implementation in rust & used gpu cores to speed up 20x!
010
Deb Sahu @debashishsahu.com · 14/05/2026
I’ve already equipped my LLMs to beat this challenge: github.com/CloakHQ/Cloa...
github.com
GitHub - CloakHQ/CloakBrowser: Stealth Chromium that passes every bot detection test. Drop-in Playwright replacement with source-level fingerprint patches. 30/30 tests passed.
Stealth Chromium that passes every bot detection test. Drop-in Playwright replacement with source-level fingerprint patches. 30/30 tests passed. - CloakHQ/CloakBrowser
000
Deb Sahu @debashishsahu.com · 13/05/2026
Voice-in voice-out: ~2m47s 😂 first turn (STT 6s + 60s prefill + 30s gen + TTS 1s). Pascal's no-tensor-cores prefill on a 16K system prompt is the bottleneck. Subsequent turns hit llama-server's prompt cache → ~30s/reply. Ampere+ would cut prefill 6x
000
Deb Sahu @debashishsahu.com · 13/05/2026
www.youtube.com/watch?v=1jhR...
youtube.com
Bambu Lab: I'm reposting your code & I dare you to sue me.
YouTube video by Louis Rossmann
000
Deb Sahu @debashishsahu.com · 13/05/2026
Gemma 4 E4B is amazing for its size, I’m waiting for upstream MTP merge to see how much I can push in this oldish hardware.
010
Deb Sahu @debashishsahu.com · 12/05/2026
Every year Bambu Labs have to create some controversy to stay relevant? Love the 3D printers and not the company! gamersnexus.net/fk-you-bambu...
gamersnexus.net
Fuck You, Bambu Lab: OrcaSlicer-BambuLab Download (with permission) | GamersNexus
News Features Fuck You, Bambu Lab: OrcaSlicer-BambuLab Download (with permission) May 12, 2026 Last Updated: 2026-05-12 With permission of the OrcaSlicer-BambuLab software, we are rehosting the DMCA-s...
010
Deb Sahu @debashishsahu.com · 12/05/2026
Built a fully-local OpenClaw AI assistant available on WhatsApp using CachyOS (Linux) box w/ an old GTX 1070 Ti (Pascal, Vulkan). Orchestration Stack: - Gemma 4 E4B for chat+vision+STT - Piper for TTS - DreamShaper 8 for image gen Voice-in/voice-out, photo descriptions, image gen: zero cloud calls.
621
Deb Sahu @debashishsahu.com · 10/05/2026
I ended up landing to this gem: trianglesplatting2.github.io/trianglespla...
trianglesplatting2.github.io
Triangle Splatting+: Differentiable Rendering with Opaque Triangles
Triangle Splatting+: Differentiable Rendering with Opaque Triangles, 2025.
000
Deb Sahu @debashishsahu.com · 10/05/2026
Someone mentioned two words “Gaussian Splat” on Friday and I went into a deep rabbit hole into this wonderful technology. Went through a bunch of papers, math/reconstruction and free open source programs. Going to try some point cloud reconstruction soon.
000
Deb Sahu @debashishsahu.com · 06/05/2026
Yay more compute = more work done
000
Deb Sahu @debashishsahu.com · 05/05/2026
Bambu Labs that everyone was warning about. They are pure evil, I hope Pawel fights back. www.youtube.com/watch?v=jIbp...
youtube.com
Bambu labs sends legal threat to orcaslicer dev over use of AGPL code 🤦
YouTube video by Louis Rossmann
000
Deb Sahu @debashishsahu.com · 02/05/2026
Love the feedback, running realistic scenarios next.
000
Deb Sahu @debashishsahu.com · 02/05/2026
WhisperKit www.forasoft.com/blog/article...
forasoft.com
iOS Speech Recognition in 2026: SpeechAnalyzer, WhisperKit & the Full On-Device Playbook
On-device vs cloud speech recognition on iOS in 2026: SpeechAnalyzer, WhisperKit, Core ML, Whisper API, Deepgram. Pipelines, WER, latency, privacy and code.
100
Deb Sahu @debashishsahu.com · 02/05/2026
mqtt2db-go v0.1.1 - addressed realistic scenarios Sustained PG slowdown stress test (toxiproxy, 10K msg/s): • moderate +300ms: drain 18s • severe +2.5s: drain 20m • outage TCP-rej: drain 8m 6/6 PASS, zero DLQ, exact conservation. Per-conn cached staging + parallel flushers: 5K → 25K rows/sec.
100
Deb Sahu @debashishsahu.com · 01/05/2026
Both fair. Stress schema was minimal (one table, unique index on dedup_key only); wider schemas with secondary indexes hit a different ceiling. And agreed: 30s clean runs prove little. Real next test is sustained PG slowdown to exercise adaptive flusher mode transitions and WAL spill. Honest gap.
100
Deb Sahu @debashishsahu.com · 01/05/2026
mqtt2db-go is the middle ground: a focused Go binary that does one thing well. Subscribe to Comqtt via shared subs, batch via pgx.CopyFrom into Postgres, with operational properties tuned for IoT telemetry at scale. No broker coupling, no schema rigidity, no licensing surprises.
000
Deb Sahu @debashishsahu.com · 01/05/2026
Why this exists: moving MQTT into Postgres has two off-the-shelf paths, and both have tradeoffs. EMQX bundles a rule engine with DB sinks but couples ingest to the broker hot path and now carries BSL licensing risk. Telegraf and NiFi work but offer little control over backpressure.
100
Deb Sahu @debashishsahu.com · 01/05/2026
Key call: the durability boundary lives at Postgres, not the WAL. MQTT messages get ack’d only after successful insert, so the at-least-once contract holds end-to-end. Throughput is bounded by COPY rate; the win is clean failure semantics. Badger smooths bursts, S3 catches poison.
000
Deb Sahu @debashishsahu.com · 01/05/2026
Architecture: Comqtt (Apache 2.0 MQTT broker) feeds mqtt2db-go, which batches via pgx.CopyFrom through PgBouncer into Postgres. Three-tier buffer absorbs bursts: in-memory ring, Badger WAL on PVC, S3 dead-letter. Ingestion decoupled from processing; backpressure propagates upward.
100
Deb Sahu @debashishsahu.com · 01/05/2026
Shipped mqtt2db-go v0.1.0: a Go service that ingests MQTT into Postgres at IoT scale, without EMQX’s BSL licensing trap or Telegraf’s schema rigidity. Stress validated at 600K messages over 30s @ 20K msg/s, zero internal loss. MIT licensed. github.com/debsahu/mqtt2db-go
github.com
GitHub - debsahu/mqtt2db-go: High-throughput Go service that subscribes to MQTT topics on a Comqtt cluster and persists messages to PostgreSQL via PgBouncer. Designed to mimic Google Pub/Sub semantics...
High-throughput Go service that subscribes to MQTT topics on a Comqtt cluster and persists messages to PostgreSQL via PgBouncer. Designed to mimic Google Pub/Sub semantics (at-least-once delivery, ...
110
Deb Sahu @debashishsahu.com · 25/04/2026
CLI tool to reduce your LLM token usage by up to 60-90% github.com/rtk-ai/rtk
github.com
GitHub - rtk-ai/rtk: CLI proxy that reduces LLM token consumption by 60-90% on common dev commands. Single Rust binary, zero dependencies
CLI proxy that reduces LLM token consumption by 60-90% on common dev commands. Single Rust binary, zero dependencies - rtk-ai/rtk
000
Deb Sahu @debashishsahu.com · 23/04/2026
No wonder Claude felt so dumb the last week! Anywho, all our usage has been reset! www.anthropic.com/engineering/...
anthropic.com
An update on recent Claude Code quality reports
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
000
Deb Sahu @debashishsahu.com · 23/04/2026
Qwen 3.6 27B (tiny model) beats opus 4.5 and can run locally on a 64GB MacBook Pro - excellent for agentic coding tasks. I’m genuinely impressed for its size huggingface.co/unsloth/Qwen...
huggingface.co
unsloth/Qwen3.6-27B-MLX-8bit · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
010
Deb Sahu @debashishsahu.com · 17/04/2026
Ohh no, another one I will be skipping. Feels like they ask for feedback - we all say that infra at Asilomar is dated and it’s just ignored. Love the people, community and science - don’t love the conference grounds.
110
Deb Sahu @debashishsahu.com · 17/04/2026
Hopefully it’s not in ASILOMAR! What’s the location?
100
Deb Sahu @debashishsahu.com · 06/04/2026
Agree!
000
Deb Sahu @debashishsahu.com · 06/04/2026
Never meant for consumer use, usually these are flashed with openwrt or something equivalent. That’s what these are used for. That being said most routers have backdoor for recovery mode, where one can do firmware dump or even flash with custom firmware.
000
Deb Sahu @debashishsahu.com · 01/04/2026
No need to hack it, leverage the open source community. All of prusa design, electronics and software is free!
000
Reposted by Deb Sahu
Magnetic Resonance News @nmr900.bsky.social · 16/03/2026
Nuclear Magnetic Resonance (NMR) Director, The University of Colorado Boulder, Colorado jobs.colorado.edu/jobs/JobDeta... #NMRjobs #NMRchat 🧲
jobs.colorado.edu
Nuclear Magnetic Resonance (NMR) Director
022
Deb Sahu @debashishsahu.com · 15/02/2026
ChatGPT
000
Deb Sahu @debashishsahu.com · 15/02/2026
DeepSeek
100
Deb Sahu @debashishsahu.com · 15/02/2026
Kimi K2.5 Thinking
100
Deb Sahu @debashishsahu.com · 15/02/2026
Perplexity
100
Deb Sahu @debashishsahu.com · 15/02/2026
Gemini
100
Deb Sahu @debashishsahu.com · 15/02/2026
Claude without thinking
100
Deb Sahu @debashishsahu.com · 15/02/2026
Claude with Extended thinking
100
Deb Sahu @debashishsahu.com · 15/02/2026
I asked bunch of LLM: Should I walk or drive to wash my car at the car wash that is only 100m away from my house?
210
Deb Sahu @debashishsahu.com · 25/01/2026
Amazing videos… side note I’ve personally used that Magritek’s earth field NMR in the background of these videos. It’s really cool to understand basic NMR/MRI concepts. However, it took me a week to figure out why I wasn’t getting any signal because the building I was in was a Faraday cage!
021
Deb Sahu @debashishsahu.com · 12/01/2026
Visited Toronto for the first time, 540 miles #EV round trip on my #Mach-E, solely relying on #Tesla #supercharging #network, averaging about $0.21/mile - not bad at all for winter driving.
020
Deb Sahu @debashishsahu.com · 18/12/2025
This hits #RHEL 9, #AlmaLinux 9, #Rocky 9, #CentOS Stream 9: all the same under the hood, all the same headache. Check your limits.conf BEFORE your next upgrade. You'll thank yourself later. #NMR #Bruker #TopSpin #ScientificComputing #LabIT
010