@torren

Today was all about building and training the LLM behind Socle Intelligence, probably one of the most ambitious parts of Socle so far.

The goal isn’t to make another AI chatbot that spits out generic advice. I’m building a system that can actually understand what’s happening inside a business by connecting customer experiences, employee performance, feedback patterns, and historical data.

For example, if a customer says their food took forever but their server was amazing, the model needs to recognize two completely different signals: a potential operational issue and an employee who deserves recognition.I’ve been working on the model’s reasoning, training it to interpret feedback in context, experimenting with different parameters, and building the infrastructure that will eventually let business owners ask questions about their own data.

The hardest part is getting the AI to distinguish between an isolated complaint and an actual emerging pattern, without making conclusions that the evidence doesn’t support.

I’m also exploring local model deployment rather than relying entirely on external AI APIs, which adds another layer of complexity.

The end goal is simple: turn thousands of individual interactions into intelligence that businesses can actually act on.Still plenty of work left, but this is where Socle starts moving beyond collecting data and toward understanding it.

Rebuilt the cat in Blender instead of stacking Roblox parts. Smooth meshes, a real open hoodie with folds and cuffs, a satchel on a leather strap, round glossy eyes and a white blaze down the face. The ears took a few tries: one version came out looking like triangles, so I brought back the older rounder ones. The first fluffy tail had bits sticking out of it and looked like a monster, so now it’s one smooth plump tail.

The blasters looked like kids’ toys, so I redesigned them as stylized sci-fi guns with real gun parts: receiver, rail, vented handguard, magazine, stock and sights, with one colour panel and a small glow at the muzzle. Each gun has its own shot sound and bolt style.

Gameplay feel was the big one. Your shots now show up instantly on your screen, and hit detection leans toward the shooter. Hits launch cats up and away instead of just nudging them. Bigger guns kick you back, and you can wall-jump and pounce (E). Bots got a skill level and a favourite gun each, keep to that gun’s range, and only shoot what they can actually see.

Also added:

• Neon Rooftops, a fourth map
• 3 spawns per team spread across different islands
• emotes that actually animate
• a hit sound that rises in pitch as the target’s knockback % climbs
• a thump and whoosh when a cat gets launched
• a slow-mo replay of your last knockout at the end of each match
• an intro for new players: a short cinematic, then Mochi, a mentor cat, teaching you each move hands-on

Next: a proper lobby.

I built OMEGA-7, a space escape room that turns Figma into a playable game. Players explore a dying station and solve puzzles before their oxygen runs out. The escape sequence involves finding a keycard, decoding an airlock PIN, and rewiring a pod. I added 3D-rendered rooms, clickable objects, camera push-ins, and hint screens. I tested the puzzle paths from start to finish to check the full escape sequence.
I also launched a browser version with real-time 3D, swinging doors, and sound effects.
It adds a seven-minute oxygen timer and a wire-dragging puzzle, linking back to the Figma prototype.

PAW MAYHEM devlog: the game has a real lobby now

The old play screen felt like a menu stuck on top of a game, so I replaced it with a walkable lobby, like Rivals on Roblox.

You spawn on a garden island in the sky, facing the duel pads. Behind you is a practice range with target cats you can knock off the back edge. On the left there’s an armory with every blaster on a pedestal (walk up and press F to equip one). On the right there’s a boutique of cats wearing featured looks, and front-right is a board showing the top trophy players. Railings and an invisible wall keep you in, and if you fall off you get put back.

Shop, Customize, Loadout and Season Pass are now small icons on the side of the screen instead of full menu pages, and PLAY sits at the bottom. Press it, pick a mode, vote for a map, and the camera flies over the arena before you drop in. Matches fill up to 16 cats with bots when there aren’t enough people.

Duels: there are 1v1 and 2v2 pads in front of spawn. Stand on a slot, and when the pad fills up there’s a 3 second countdown, then you get your own duel on one of three new maps: Sky Bridge, Rooftop Duel and Crystal Ring. First to 5 knockouts wins trophies. If nobody joins within 15 seconds, a bot takes the empty slot. Bot duels don’t give trophies, so you can’t farm them.

You can also change your loadout mid-match now. Press B and pick a blaster. If you’re in the first 5 seconds after spawning it goes straight into your paws, otherwise you get it when you respawn.

The annoying part: the first time I pressed Play, I just fell forever. The lobby is far from everything else, and Roblox streaming hadn’t sent the floor to my game yet, so my cat fell through, got teleported back, and fell again. I turned streaming off and made spawning wait until there’s ground under you.

spent most of this session working on one of the bigger systems behind socle: the analytics/intelligence engine.

the goal is basically to make sure socle does more than just collect feedback and throw a bunch of charts at a business. if someone scans a Snap, leaves feedback, recognizes an employee, saves a place, visits somewhere, or interacts with socle in another way, that creates a signal. the harder part is figuring out what those signals actually mean when you start connecting them together.

so i started building out the system that takes those interactions and turns them into actual insights and suggestions.

for example, instead of just saying “service mentions went down,” socle should be able to notice the pattern, look at what guests are actually saying, which types of guests are experiencing it, whether it is happening repeatedly or is just an outlier, and then suggest what the business could actually do about it.

i also spent a lot of time on the explanation side of it. i really don’t want the dashboard to randomly spit out recommendations with no idea where they came from. every insight needs to have evidence behind it, so businesses can see why socle thinks something is happening, what signals contributed to it, and how confident the system is.

if it gets something wrong, they should also be able to report that so we can keep improving the logic.the engine connects things like written feedback, selected responses, recognitions, interaction history, guest taste/context, repeated patterns, and changes over time. when multiple signals point toward the same thing, the system can treat that differently from one random interaction instead of pretending every piece of data means the same amount.

i also worked on how this splits between the business and employee dashboards. managers need broader operational suggestions and patterns across the business, while an employee should get useful feedback about what they are doing well and what they could improve without just getting a weird analytics graph thrown at them.a big rule while building this has been no fake intelligence.

if there is not enough evidence to make a useful conclusion yet, socle should say that instead of inventing one. right now a lot of the engine is deterministic/rules-based so i can actually test why every result appears. later i want to add an AI layer on top so businesses can ask questions directly and the system can reason across all of this context, but i want the underlying data and evidence system to be solid first.this is probably one of the first updates where the two sides of socle really start connecting together.

the guest-facing product creates the interactions, Snap gives us another real-world interaction point, and the intelligence engine turns all of that into something a business can actually use.still a lot to build and test, but this is starting to feel a lot less like “analytics” and a lot more like the intelligence layer i originally wanted socle to become.

spent around 11 hours on socle today, which is definitely over the devlog limit because i completely forgot to log before getting locked in 😭

this was one of those sessions where i fixed one tiny thing, noticed three more things that felt off, fixed those, then somehow found even more. i went through a bunch of screens and flows trying to look at everything like someone using socle for the first time instead of someone who already knows where every button is. i cleaned up spacing, navigation, loading states, interactions, weird edge cases, and a lot of smaller details that were not huge on their own but made parts of the app feel unfinished.

there was not one massive new feature this time, but the app feels much more solid now. i think we are getting to the point where the remaining work is mostly about removing those little moments of friction that make a product feel rough, even if users cannot always explain exactly why.

we were also hoping to ship this version, but we ran into some repo issues, so we are trying to figure out a cleaner solution before releasing it. hopefully we can sort that out soon and get this version out.

spent around 9 hours going through socle and just fixing anything that still felt annoying, confusing, inconsistent, or unfinished.

at this point a lot of the bigger stuff is already built, so i focused more on how everything actually feels to use. i went through different screens, flows, buttons, navigation, loading states, and smaller interactions and changed a bunch of things that were still kind of rough.

a lot of the changes were pretty small by themselves, but together they made the app feel way smoother. i’m trying to get socle to the point where you don’t really have to think about the ui at all because everything just works how you expect it to.

i also tried looking at everything more like a new user instead of the person who built it, which helped me catch a lot of stuff i was probably just used to seeing before.

next i’m gonna keep cleaning up anything that still feels off, but i also want to start moving more time back into bigger features and systems again. then… SHIP IT!!

spent around 8 hours working mostly on the taste system and making it easier to actually test.

the biggest thing i changed was visit system. since we don’t have the full location tracking system built yet, i replaced the old visit flow with a simple button for now. now we can actually test how visits affect someone’s taste without waiting for the full tracking system to be done.

i also changed some of the math behind taste, mostly how different actions are weighted and how repeated activity is counted so one place can’t mess up someone’s entire profile.

the screenshot below isn’t the exact equation or exact part i changed. i’m trying to be careful with how much of the scoring system i show. if you look at it and have no idea what’s going on, that’s kinda on purpose. i don’t want people to be able to reverse engineer it later or find ways to manipulate it if there are any loopholes we haven’t caught yet.

next, i’m gonna keep testing the weights, keep doing ui/ux and qol polishes, and just keep fixing anything that still feels off before we bring in more testers.

spent today building paw mayhem, a physics-based roblox multiplayer brawler where you play as a customizable cat and shoot enemies off floating islands. built the full game from scratch using rojo — server-authoritative knockback physics, 3 procedural maps (sky islands, volcano, toybox), and 8 weapons each with a unique silhouette. wrote all the ui in code: main menu with an animated hero scene, in-game hud, customize, shop, loadout, leaderboard, quests, and a victory screen with confetti. also set up xp/level progression, a coin economy, daily quests, and mid-match power-ups. zero toolbox models, every part, icon, and map decoration is built from roblox primitives.

spent around 9 hours doing another huge UI/UX polish pass on socle.

at this point most of the main stuff is already there, so now i’m mainly finding all the smaller things that make the app feel unfinished. stuff like weird spacing, inconsistent layouts, interactions that feel a little off, small bugs, or certain parts of the app just not feeling as polished as everything else.

for most of this session i was basically just going through the app over and over and fixing anything that felt wrong. cleaned up a bunch of layouts, fixed smaller UI/UX issues, changed some interactions, and reworked parts of the app that still felt kind of rough.

i also added more locations and more content to existing locations so there’s actually more to explore when people start testing. that’s becoming more important now because as we bring in more testers, i don’t want them opening the app and running out of stuff to do after like 5 minutes.

didn’t really add one massive new feature this time, but the app feels way better than it did before this session.

everything is starting to feel a lot more complete now and we’re getting way closer to being ready for a much bigger testing wave.

spent around 9 hours doing another huge UI/UX polish pass on socle.

at this point most of the main stuff is already there, so now i’m mainly finding all the smaller things that make the app feel unfinished. stuff like weird spacing, inconsistent layouts, interactions that feel a little off, small bugs, or certain parts of the app just not feeling as polished as everything else.

for most of this session i was basically just going through the app over and over and fixing anything that felt wrong. cleaned up a bunch of layouts, fixed smaller UI/UX issues, changed some interactions, and reworked parts of the app that still felt kind of rough.

i also added more locations and more content to existing locations so there’s actually more to explore when people start testing. that’s becoming more important now because as we bring in more testers, i don’t want them opening the app and running out of stuff to do after like 5 minutes.

didn’t really add one massive new feature this time, but the app feels way better than it did before this session.

everything is starting to feel a lot more complete now and we’re getting way closer to being ready for a much bigger testing wave.

this session was all about making starboy feel real before the parts even show up. the website now opens with an intro where he hops onto a dark stage, a spotlight flickers on and finds him, and when you hit “leap into the experience” he jumps straight at the camera and his screen opens up onto the page. i wrote a renderer from scratch so his body reads as actual polished chrome instead of flat grey, and gave him new glowing mint eyes. i also moved the keyring off the outside of the star and into a hole through its point, rebuilt the 3d files around it, and re-rendered everything in black chrome with a steel ring through it. every link and heading on the site now reacts when you hover, with letters that roll and ripple as your cursor moves over them. next up is building the real one once the parts arrive.

this session was all about making starboy feel real before the parts even show up. the website now opens with an intro where he hops onto a dark stage, a spotlight flickers on and finds him, and when you hit “leap into the experience” he jumps straight at the camera and his screen opens up onto the page. i wrote a renderer from scratch so his body reads as actual polished chrome instead of flat grey, and gave him new glowing mint eyes. i also moved the keyring off the outside of the star and into a hole through its point, rebuilt the 3d files around it, and re-rendered everything in black chrome with a steel ring through it. every link and heading on the site now reacts when you hover, with letters that roll and ripple as your cursor moves over them. next up is building the real one once the parts arrive.

spent around 6 hours doing a big UI/UX cleanup across socle.

i went through a bunch of different screens and just fixed a lot of the smaller things that felt off. stuff like weird spacing, layout issues, awkward interactions, inconsistent components, and parts of the app that still felt a little unfinished.

i also cleaned up some of the navigation and made things easier to understand, especially for people using socle for the first time. most of the changes were pretty small on their own, but together they made the app feel way more polished and consistent.

the main goal was just getting everything into a much cleaner and more stable state before we bring in a bigger wave of testers. i’d rather fix all the obvious stuff now so when people test it, their feedback is actually about the product and not random issues we already knew were there.

parts aren’t here yet so this one was all software. the firmware compiles for the first time, and i wrote a hardware test that checks the screen, backlight and every sensor one by one so i can catch a bad wire before closing the shell. i added a tuning console so i can make him shiver, play any effect or change how sensitive he is over usb without reflashing. then i wrote tests for his mood logic, and they caught 3 bugs that would’ve only shown up on the real star: shaking froze his eyes, hanging on a belt loop counted as tilted forever so he’d never sleep, and my battery saver made him deaf while asleep. the website got a sensor sandbox where you control his temperature, noise, tilt and shake, plus you can tap any of the 400 eye looks to get the exact seed for a real star.

parts aren’t here yet so this one was all software. the firmware compiles for the first time, and i wrote a hardware test that checks the screen, backlight and every sensor one by one so i can catch a bad wire before closing the shell. i added a tuning console so i can make him shiver, play any effect or change how sensitive he is over usb without reflashing. then i wrote tests for his mood logic, and they caught 3 bugs that would’ve only shown up on the real star: shaking froze his eyes, hanging on a belt loop counted as tilted forever so he’d never sleep, and my battery saver made him deaf while asleep. the website got a sensor sandbox where you control his temperature, noise, tilt and shake, plus you can tap any of the 400 eye looks to get the exact seed for a real star.

i’m building a tiny wearable pet from scratch: a chrome star with a round screen for a face that clips onto your belt loop. he reacts to you, so shake him and he gets dizzy and mad, take him somewhere cold and he shivers, and yell near him and he gets nervous. every star rolls his own eyes on first boot, with 4 shapes and 100 colorways for 400 looks, and some only show up 0.04% of the time. i designed the 20.5mm shell in openscad with a stepped pocket so the 37.5mm screen board actually fits, hand-wired everything to a seeed xiao esp32c3, and finished the eye firmware with 20 rare effects. you can play with him live at starboy-diy.nadellasripad11.workers.dev, and next up the parts arrive so i can print and chrome the shell and wake him up for real 👀

this update was mainly focused on UI and UX. i used codex to help with some of the code implementation while i went through different parts of the app and tested everything.

while testing, i found a few bugs in one of the features i had already built, so i went back and fixed those too. after the changes, the feature was working properly again and the overall experience felt a lot cleaner.

next, i want to keep going through the app and completely get rid of any broken UI, bad UX, or annoying little issues. the goal is basically to get the QoL as close to perfect as possible before more people start testing it.

star shaped keychain with a round display eye that reacts to shaking, cold, noise and gets sleepy when left alone. runs on an esp32-c3 with an accelerometer, temp sensor and mic, and every unit generates its own eye colours and pupil style from a seed so no two come out the same. the shell took a few tries to get right, first a rounded blob then long thin spikes, before landing on a fat middle with short points. caught a display pocket that was 4mm too small (every listing quotes the glass size, not the board) and the back engraving cutting into the electronics cavity, then added checks in the cad so those can’t sneak back in. stls are exported and ready.

this update was mainly focused on UI and UX. i used codex to help with some of the code implementation while i went through different parts of the app and tested everything.

while testing, i found a few bugs in one of the features i had already built, so i went back and fixed those too. after the changes, the feature was working properly again and the overall experience felt a lot cleaner.

next, i want to keep going through the app and completely get rid of any broken UI, bad UX, or annoying little issues. the goal is basically to get the QoL as close to perfect as possible before more people start testing it.