Amrish (@zmrishh) header
Amrish (@zmrishh) profile picture

22 · co-founder & ceo, https://xitter.gq/t.co/8u1mu77uBx @TheResidency 26' · @at_Altspace Fellow · incubated at @iTNTHub

Joined November 2022
Photo by Amrish (@zmrishh)Photo by Amrish (@zmrishh)Photo by Amrish (@zmrishh)Photo by Amrish (@zmrishh)Photo by Amrish (@zmrishh)Photo by Amrish (@zmrishh)Photo by Amrish (@zmrishh)Photo by Amrish (@zmrishh)Photo by Amrish (@zmrishh)Photo by Amrish (@zmrishh)
Pinned Tweet
made my design/dev resource shelf public. components, ui inspo, motion, shaders, weird visual tools, and sites i actually use when building. stack.zmrish.com
1
4
1,328
why does Tiger Muthuvel Pandian lowkey have the backstory of mike ehrmantraut?
23
Agaram 2.5B is out. It's a Tamil-first language model we trained from scratch at @MuraiLabsHQ , our small non-profit lab. As far as we know, it's the first Tamil model trained from scratch that can call tools. It chats in Tamil, Tanglish and English, translates, and can call tools. We're releasing the base model, chat model and GGUF files under Apache-2.0. You can download it and run it offline on your own laptop. The first chat model had a four-day public preview. Seeing where it fell short with real users helped us decide what to work on. We rebuilt the fine-tuning around what we saw, using 222,662 records. Thank you to everyone who spent time with that early version. The video walks through the model. You can try it here: hf.co/spaces/mechramc/agaram… நன்றி 🙏
113
422
35
2,277
90,548
Amrish retweeted
Replying to @theresidency
@theresidency demo day. apparently i’ve been in bangalore long enough to have one of these now. pitched @KenesisLabs on october 22, at @residencyBLR founders track showcase, powered by Lightspeed India a few updates since the “got backed by the residency” post: >finished our pilot with SPP India >two LOIs signed >a third under review still plenty to figure out. but good to have more to talk about than what we’re planning to build. we’re building ai agents for factory floors. cameras already see what’s happening. we’re making them useful while something can still be done about it. the last few weeks have been factory follow-ups, building, and figuring out how to explain all of this in a few minutes on stage. turns out the last part needs its own preparation. met some really cool founders here. good conversations, a lot of laughs, and people i’m glad i got to spend these weeks with. thank you @AadityaYuvraj , @NimishaChanda , M S Mihir and @rishabsg17 for backing us and putting this together. @daniel_das07 , @dheekshithoff, Aswin. a lot behind what went on those slides. now back to getting this into more factories.
3
5
1
13
408
Amrish retweeted
trying to build something interesting for the london startup scene. inspired by @Londonmaxxing, @fooming86, @boardyai and @nomadtableapp (@jayvraavi). the idea is to help founders, builders and operators find their people, make connections and get things done. still early. would love feedback from people actually building in london, uk should i launch a test version? 👀
8
4
1
17
1,100
Amrish retweeted
We're welcoming the best VC funds in India to come join us for @residencyBLR's Demo Day on 22nd Oct Who's on stage: >deeptech founders building in aerospace, biotech, energy etc >consumer founders who've done $2M ARR & $200K pre-revenue in the US Link: luma.com/4qcjcvp4?tk=0lb2pk
2
9
233
3 months ago, we were just building. 2 signed LOIs, a live factory pilot, and a lot more to figure out. back to building kenesis 🫡
3 months after a @ChatGPT Pro grant, @zmrishh's team at @kenesislabs has: • 2 signed LOIs • A live factory pilot • A @nasscomai Gamechangers Top 5 shortlist They're using factory CCTV for safety and process insights. Codex helped them build faster. Congrats to the team!
6
153
big thanks to @SiddhantSadangi for giving me 3 months of chatgpt pro we’ve since signed 2 LOIs, with a third in review. Codex has helped our small team build so much faster at Kenesis. really appreciate the support, man 🙏
1
110
also sent you a detailed progress update in DMs. do check it when you get a chance 🙏
25
the final demoday is here! > one already has $2M+ in sales. > one is working on autonomous aircrafts > one prev. ran a million-dollar company, and is already at $200k+ with this one. > all of them are looking at a multi-billion-dollar outcome showcasing 8 teams building across biotech, consumertech, hardware, physical AI and more! watch them demo live on 22nd Oct at @residencyBLR final demo day in bangalore. powered by @lightspeedindia | banking partner @IDFCFIRSTBank. limited seats. link in the comments. come by!
5
8
6
33
2,752
mankind went from inventing the wheel to putting people on the moon, only to create WhatsApp for Windows. the worst fucking app in the history of human civilization. its only job is to send messages and even that feels like asking too much.
1
66
$870 million raised at a $7.5 billion valuation led by @a16z
launched a model accidentally served trillions of tokens a day 29.4% of the fortune 500 showed up raised a really big series A from @a16z it's been 3 weeks we would like to sleep now
4
306
launched a model accidentally served trillions of tokens a day 29.4% of the fortune 500 showed up raised a really big series A from @a16z it's been 3 weeks we would like to sleep now
563
308
287
8,025
1,336,538
I'll give ₹1 to everyone who reposts this
Introducing Arcanoris. Your customers stopped googling. They ask ChatGPT what to buy. If AI doesn't name you, you don't exist. Arcanoris shows what AI says about your brand, who it recommends instead, and exactly what to fix. Free audit, no card, ~2 min → arcanoris.in
72
Introducing Arcanoris. Your customers stopped googling. They ask ChatGPT what to buy. If AI doesn't name you, you don't exist. Arcanoris shows what AI says about your brand, who it recommends instead, and exactly what to fix. Free audit, no card, ~2 min → arcanoris.in
2
2
1
9
565
How it works: → Paste your site → We ask AI the questions your buyers ask → See your score, every answer, and who beats you → Get a ranked fix list Free checks ChatGPT. Plus adds Claude, Perplexity, Grok and Mistral.
1
57
demo day 4 at @residencyBLR founders track this week @joinef india team joined our weekly demo sync and residents presented across biotech, consumer hardware and defence, and everyone walked away with valuable feedback and a clear list of what to fix before next week. thank you @iampbiswal and @BLRxZo team for making it happen. onwards 🚀
1
9
1
24
1,164
W Anthropic
The Claude Startups program is expanding to more founders. Members can get a year of Claude Team, API credits, special offers on the tools a startup runs on, and office hours with Anthropic’s Applied AI team.
4
17
1,266
Goated wallpaper
Liftoff of Starship's first orbital flight
110
saw this and had to try it for Kenesis. here’s my version. pretty wild what a good reference and some direction can do.
do NOT asl Opus 5.5 to make videos from scratch. good quality videos still come from the reference, taste + direction you give it. without that, it just defaults to the same generic AI slop everyone is already posting. left is @claudeai’s original. right is my version rebuilt for fun. same pacing, same cuts, same movement. took around 30 minutes and cost $0. give it great work to study, then let it handle the execution. prompt below ↓ <inputs> Ask me for: three words for the headline build (one per act), the product line for the first end card, a wordmark name, a simple mark (or draw one), a palette mood for the horizon, a source of real photographs of round objects, and the music. If I skip any, use these defaults: words "Lorem's" → "ipsum to" → "dolorem"; product line "Lorem 5.5"; wordmark "Ipsum"; mark = a circle ring with its lower half filled solid; horizon = warm amber dusk over a black planet; photographs = CC0 museum objects from the Cleveland Museum of Art open-access API (plates, coins, mirrors, shields, roundels, stained glass, carved discs); music = the reference track. </inputs> <direction> A 20 second 16:9 brand film, 1920x1080, 25 fps, frames f1–f500. Concept: the edge of the world. Every shot is the curved top edge of a real round object filling the bottom of frame like a planet's horizon, and the cut lands on a new object every 5–31 frames. Motion law: every shot is a STILL photograph with at most a tiny camera drift. All energy comes from the cuts. No in-shot animation except where listed. The word on the arc never animates; it only changes size and texture at each cut. Look: macro still-life photography: real light, real wear, shallow focus where stated, film grain on every frame (monochrome luminance, alive every frame). Type: one classic serif (Georgia regular) for every word, white or ink per shot, sitting just above the arc. Banned: crossfades, eased moves between shots, procedural noise or sine-wave "textures", rainbow or spectral gradients, purple, violet or magenta anywhere, recolouring the reference's own images, any real brand's mark or words, Math.random. </direction> <structure> Coordinates are px on 1920x1080. Hard cuts at f32, 58, 84, 95, 101, 107, 120, 135, 148, 162, 177, 190, 204, 219, 227, 237, 245, 256, 268, 276, 286, 294, 302, 313, 320, 326, 334, 342, 349, 357, 363, 370, 377, 384, 390, 395, 400: keep these exact frames (37 cuts, verified by a frame-difference scan). Every object is placed so that its rim matches that shot's arc: a circle through (0, yL), (960, yC), (1920, yR). Its crest sits at x960. Scale = arc radius / the object's own rim radius, capped at 1.8× so the photo is never visibly upscaled. S1 f1–31, horizon open: A planet limb with radius 6881 px, centred x960, crest y≈518. Above it a glow band ~330 px thick (warm amber, exponential falloff), a thin bright line 8 px above the limb, and black planet below. Camera push 1.000 → 1.026 with x drift +28 px and y +4 px, linear across the shot. S2 f32–57: dark ground; a terracotta dish rim, focus falling off downward. Camera pan x +32 px across the shot. Overscan the plate by (1 + 2|dx|/1920) so no edge ever shows. S3 f58–83: a marbled bowl rim on black, softened 10 px. STATIC (the reference moves less than 4 px). S4 f84–94: a red-figure plate on a pale grey sweep, contact shadow 50%. S5 f95–100: slipware dish on sage green. S6 f101–106: a blue lustre bowl on near-black, rim light 10 px. S7 f107–119: blue-and-white porcelain on navy paper; arc (706, 542, 706). S8 f120–134: carved celadon on black; camera push 1.000 → 1.008. S9 f135–147: celadon ring plate on cream paper. S10 f148–161: gold-and-paste brooch (scalloped edge) on dusty pink. S11 f162–176: gold repoussé plaque on black, warm rim light. S12 f177–189: patinated bronze mirror on charcoal. S13 f190–203: plain celadon dish on tan paper. S14 f204–218: the only animated shot. A Muybridge bird-in-flight plate (public domain) advances one cell per frame on cream paper, with a hand-painted ochre hoop (radius 1250, centre (960,1700), wobbled edge) as the ground line. TEXT BEGINS. Each shot gives: word, word width in px, centre x, baseline y. Size the font so the reference word's width matches. S15 f219–226: Lorem's 225 / 952 / 509. Inlaid brass tray on smeared dark wood, blurred 5 px, exposure 0.55. Word smeared horizontally 10 px, slightly rough. S16 f227–236: Lorem's 245 / 957 / 508. Cracked clay on black. Word blur 0.6. S17 f237–244: Lorem's 270 / 960 / 519. Imari plate on orange paper; arc (560, 552, 562). Word rough-edged. S18 f245–255: Lorem's 292 / 954 / 518. Bronze mirror back on black. S19 f256–267: Lorem's 325 / 952 / 524. Stained-glass roundel on black, warm rim. Word blur 1.4, glow 0.35. S20 f268–275: ipsum to 370 / 955 / 510. A dark lacquer box backlit on burnt orange, blurred 9 px; arc (660, 600, 660). Word 80% opacity, blur 2.2. S21 f276–285: ipsum to 396 / 955 / 510. Carved black lacquer silhouette on brown dusk; arc (730, 650, 780). Word 55% opacity. S22 f286–293: ipsum to 419 / 957 / 511. Bouclé wool over ribbed corduroy, frayed edge. Word textured by the corduroy. S23 f294–301: ipsum to 445 / 957 / 511. Silver-framed enamel rim on mustard, soft 6 px; arc (760, 610, 760). S24 f302–312: ipsum to 450 / 963 / 495. Pierced creamware basket on black. S25 f313–319: dolorem 580 / 960 / 503. Hammered bronze on brick red. S26 f320–325: dolorem 590 / 963 / 518. Blue-and-white bowl on a dark peach ground; the white word must have contrast. S27 f326–333: dolorem 625 / 952 / 523. A gold coin rim on black, focus falloff; arc (700, 600, 700). The rim stays fully below the word. S28 f334–341: dolorem 625 / 957 / 508. Silver repoussé flowers on black. S29 f342–348: dolorem 675 / 952 / 523. Beaded silver dish on cracked plaster; arc (696, 576, 650). A diagonal cast shadow over the right ~45% of the frame darkens both the plate and the word. S30 f349–356: dolorem 690 / 960 / 523. A gold seal rotated 90° on sage paper. S31 f357–362: dolorem 700 / 956 / 513. Ink word on a Delft plate over off-white. S32 f363–369: dolorem 765 / 962 / 518. Ink word on an embroidered textile roundel. S33 f370–376: dolorem 865 / 957 / 545. A polished mirror on dark ember; arc (560, 440, 560). S34 f377–383: dolorem 800 / 965 / 518. Ink word on a fretwork bronze cover over off-white. S35 f384–389: dolorem 820 / 965 / 523. Ink word on a carved stone disc over parchment; arc (576, 410, 576). S36 f390–394: dolorem 830 / 965 / 518. Silver repoussé blurred 7 px on black; arc (760, 560, 760). Word blur 3.2. S37 f395–399: dolorem 825 / 962 / 518. Carved stone glyph on black, warm rim. S38 f400–500, end card: The S1 horizon world, with the limb low in frame (glow rising from y≈850). "Lorem 5.5" centred at (960,540), sized so "Opus 5.5" would be 377 px wide plus 9%, Georgia with lining digits. Camera drift +100 px over f400–455. At f433 hard-swap to mark + "Ipsum" at about the same size and centre. f456: the camera starts tilting up. Per-frame upward velocity f457–500 (px): 1.25, 1.75, 2, 2.25, 2.75, 3.25, 3.5, 4, 4.5, 5.25, 5.5, 6, 6.5, 7.25, 8, 8.5, 9.25, 10.25, 10.75, 11.75, 12.5, 13.75, 15, 16, 17.75, 19, 20.5, 22.5, 24.25, 26.25, 28.5, 30.5, 32, 34.5, 37.5, 40.75, 44, 46.25, 47.5, 46.25, 42.25, 35.5, 27.5, 15.25 (total 810). The logo rides the world and blurs vertically by 0.4 × velocity, interpolated continuously. f500 must equal the f1 view, so the film loops seamlessly. </structure> <build> 1. One renderer that draws each frame as a pure function of the frame number: seed(shot start) for every random value, no timers. 2. Plates: download museum print JPEGs, model the grey sweep with a 2-D polynomial fitted to the border, threshold, keep the single connected object, close and open with 9 px, choke 3 px, feather 1.8 px. Fit a circle to the top rim. Pad each cutout 900 px with edge-extend so no photo border ever reaches the frame. Never apply a contrast curve to the photographs: museum JPEGs posterize. 3. Cache each still shot's plate once and add per-frame grain on top, so nothing freezes. 4. Before rendering, scan the shot table for one-frame outliers and any bad first or last value. Block the render if any exist. 5. Encode H.264 yuv420p, 25 fps, CRF 15, mux the reference audio. Count the frames: 500. Also build a synced side-by-side (REF | REMAKE). </build> <gotchas> - Fitted arcs fail on blurry, dark or line-drawing shots. Hand-set those with the three-point circle. - Upscaling a museum photo past ~1.8× looks mushy at 1080p. Pick a bigger object instead. - PIL floodfill silently does nothing on a numpy-wrapped image; .copy() it first. - Particles must sit behind the object and the word, never on top. - A bright rim crossing the word kills legibility. Keep every rim below the baseline. - The reference shot S3 looks like a pan but is static. Measure before you animate. - Run a motion critic and a design critic (separate agents, default stance reject) at full resolution. The design critic judges "as expensive as the reference, frame vs frame". Loop until both score 8 or more. </gotchas> <start> Ask me for the inputs. Then show me 4 stills (f2 the horizon, f170 the gold plaque, f330 a text shot, f440 the end card) side by side with the reference before you render the full film. </start>
4
615
Needed this
1
51
Saw this at the beginning of a yt video
10