build beautiful things

maglev (pinned)

where we think different about manufacturing.

visit ↗x linkedin

joint tuner @ fireworks: deploy

first place overall a gold medal on a red ribbon, with a little crying cat poking its face, paws and feet out of it

sponsors: fireworks, microsoft, nvidia

sissi on stage holding up a giant check next to the host with a microphone: fireworks: deploy, first place, $10,000, pay to the order of joint tuner, november 02, 2026, san francisco, with the fireworks, microsoft, nvidia and hacker//squad logos along the bottom — first place, $10,000, fireworks: deploy, san francisco.

the live demo on youtube, "sissi wang - joint tuner | fireworks: deploy with microsoft & nvidia", 5 minutes 4 seconds, from hackersquad. the still shows joint tuner's title slide (github.com/sissississi-013/joint-tuner, trained on fireworks, benchmarked vs gpt-6.1 terra) beside the fireworks: deploy card and a small camera view of sissi presenting at a lectern — the live demo on stage, on hackersquad's youtube (5:04).

the demo, with my voice (1:49).

a race on one wrist joint. the classic optimizer needs 493 physical tests, gpt-6.1 terra gets it on the 1st after 23 s of thinking, joint tuner on the 5th, in about a second each.

a motor + gearbox. the orange output shaft runs joint tuner's gains in mujoco and settles inside the spec. real time.

a robot arm's elbow, driven by joint tuner's gains, meeting its spec. 37x speed.

a humanoid ankle that can't balance on its own. joint tuner's gains hold it steady. real time.

a drone's roll axis, with the lag of real rotors, settling on joint tuner's gains. 2x speed.

diagram: a fine-tuned model on fireworks proposes 8 gain sets, a control-theory judge simulates them and throws out bad ones before any hardware moves, nelder-mead repairs if none pass, and the passing gains go to the real joint. trained once on 7,500 simulator-verified designs for $5.11 — how it works: a small model fine-tuned on fireworks proposes 8 gain sets, a simulator throws out the bad ones before any hardware moves, and an optimizer repairs the rest if needed.

bar chart of tasks solved out of 100 held-out controleval tasks: joint tuner with repair 93, optimizer with 600 tests 93, joint tuner 8 samples plus repair 90, kimi k3 thinking 90, gpt-6.1 terra 89, down to small base models at 8 to 11 — on 100 held-out tasks, joint tuner solves 90, the same as gpt-6.1 terra (89) and kimi k3 thinking (90), at a fraction of the model time.

bar chart: fine-tuning the 35b model on fireworks for $5.11 took its first answer from 20 to 50 tasks solved, and best of 8 from 53 to 81 — fine-tuning on fireworks for $5.11 took the 35b model from 20 to 50 right on its first answer, and from 53 to 81 with best of 8.

loop diagram: propose, test, log, retrain. each step test becomes a labeled example, and feeding them back into fireworks training cuts the tests the next joint needs — every tuning job leaves labeled tests behind; feeding them back into training means the next joint needs fewer tests.

the event cover: "fireworks: deploy", november 02, san francisco, with the fireworks, microsoft and nvidia logos along the bottom — fireworks: deploy, nov 2 in san francisco, with fireworks, microsoft and nvidia.

hackersquad github demo video linkedin

nov 2026

molten: an a16z la techweek spatial hackathon @ xtribe, with iana

supported by a16z techweek student grant

sponsors: elevenlabs, sidefx, rokoko, roboflow, miro, assemblyai, niantic spatial, kaedim, datafeel

a person walking past a big screen reading molten tribe in a brick-and-steel warehouse — xtribe, inglewood.

three organizers on a low stage beside a schedule slide, the room listening — opening molten.

an organizer speaking into a mic on stage while sissi stands beside her with a laptop — on stage.

three organizers talking in the warehouse, one pointing across the room — figuring out where everything goes.

a wide shot of a high-ceilinged warehouse full of desks of people hacking — a warehouse full of builders.

close-up of hands plugging jumper wires into a breadboard next to a dev board — wires and breadboards.

a hand resting on a dark tactile tablet with a braille display — a refreshable braille tablet.

a hacker lowering a vision pro headset over his eyes — headsets on.

building in the headset.

a hacker in a white vr headset at a busy table — testing in vr.

three hackers laughing at their table, one throwing a peace sign — good company.

a datafeel engineer bent over a demo with a hacker, a green plush toy on the table — datafeel helping a team.

salad bowls and food laid out on a long counter while people serve themselves — lunch.

a speaker with a mic in front of a houdini production slide — a talk from sidefx.

rows of hackers in orange chairs turned toward a talk — listening in.

motion capture, live.

a builder presenting a game on a big screen to a seated crowd — demos.

one builder wearing a vr headset while his teammate explains to the audience — a vr demo on stage.

a team of three presenting from armchairs in front of the crowd — another team up.

an organizer speaking into a mic while another claps beside her — closing.

the winning team holding their prize next to a slide reading overall winning team, crowdcast — awards: crowdcast, overall winner.

the winning teams lined up together in front of the screen — the winners.

visit ↗

oct 2026

interface: hack in between: an a16z sf techweek hackathon @ founders inc, with iana

my first time hosting a hackathon!

supported by a16z techweek student grant

sponsors: elevenlabs, assemblyai, bright data, tenstorrent, sensory, kaedim, panic, pupil labs, datafeel, genki instruments, eyegaze, dot inc., infineon, saleae, digilent, mcneel, niantic spatial

the long white pier building at fort mason over blue bay water, seen through a window frame — fort mason, the morning of.

an empty sunlit loft with soft green armchairs and potted trees around a big screen reading interface — the stage before doors opened.

pupil labs eye-tracking glasses laid out on a round wooden table with a sponsor card — hardware tables: pupil labs eye trackers.

stacked boxes of digilent analog discovery pro and zybo boards on a curved wooden table — digilent boards, ready to borrow.

iana standing with a laptop and sissi sitting in a green armchair next to the interface screen — me and iana, just before kickoff.

iana and sissi on either side of the interface screen, speaking to the room — opening the hackathon.

iana smiling as she lifts food out of a foil catering tray with tongs — iana serving lunch.

sissi in a green armchair with a mic, the screen behind her reading bambu lab 3d printers — me introducing the bambu lab 3d printers.

a big bright loft full of tables of people hacking on laptops — a full room by afternoon.

two hackers leaning into a laptop screen, an orange water bottle in front — heads down.

five hackers grinning at the camera around a table of laptops — a team mid-build.

a hacker talking to the datafeel team at their black-draped table — trying datafeel's haptics.

a hacker trying a hardware demo at a high table while a sponsor explains it — sponsor demos all day.

a builder presenting on stage, the screen reading talk to a robot — demos: talk to a robot.

hackers seated at tables turned toward the stage, listening — everyone watching demos.

a builder presenting next to a bright yellow slide about a playdate app — demos: a playdate for the walk home.

sissi in an armchair with a mic, the screen reading thank you for bringing your ideas to life and helping each other along the way — me, saying thank you.

a winner holding a prize next to a slide reading spidey sense — awards: spidey sense.

two winners standing by a slide reading haptic ride — awards: haptic ride.

a group photo of hackers and organizers lined up in front of the interface screen — everyone who stayed to the end.

visit ↗

oct 2026

voice: a model that writes like me @ yc own your intelligence hackathon, with iana

first place, river ai side quest a gold medal on a red ribbon, with a little crying cat poking its face, paws and feet out of it

sponsors: river ai, gbrain, memorable, qm, superset, ufo

the loom demo, "training ai to write in my voice", 1 minute 45 seconds. the still shows voice's readme: the same sentences rewritten by the base qwen model and by sissi's fine-tuned lora, which writes lowercase and casual like her — the demo on loom (1:45): it learns from my own messages, then a little bubble offers 3 replies in my voice in any text box.

a full room of hackers at the yc own your intelligence hackathon, sitting in rows facing a stage, under round paper lamps — the room at yc (photo: compiled)

a group of hackers standing together on the stage at the end of the yc own your intelligence hackathon, slides behind them — everyone on stage at the end (photo: compiled)

github demo video on loom x read more :) on compiled

sep 2026

bobot @ founders inc, silly robots contest, with iana

it speaks minionese, sings songs, loves bananas, recognizes people and hands you a banana if it wants to make friends with you.

grand (silliest) prize a gold medal on a red ribbon, with a little crying cat poking its face, paws and feet out of it

sissi and iana smiling at a table, holding bobot: a fuzzy brown minion-ish robot head with big blue goggle eyes, wires trailing to the table — me and iana with bobot.

bobot on a hackathon table: a fuzzy brown bear-like head with big yellow-and-blue goggle eyes on a white robot base, a white robot arm with a gripper beside it, and a bunch of bananas — bobot at home on the table: goggle eyes, a robot arm, and its bananas.

five people on stage in front of a bright pink and green silly robots contest screen that reads sept 10, fort mason; sissi and iana hold an orange robot hand between them — on stage at the silly robots contest, fort mason, sept 10.

github x

sep 2026

offshore: an email-native ai ops agent for offshore oil platform maintenance @ openclaw hackathon

second place a gold medal on a red ribbon, with a little crying cat poking its face, paws and feet out of it

sponsors: eragon, nozomio, agentmail

the in-person demo (4:38): a quick improvised pitch in the first round, streamed from rednote.

the loom demo, "offshore, 26 april 2026", 5 minutes 10 seconds. the still shows offshore's dark field report screen saying "ready when you are", with a start call button and a side panel for logs and activity — the demo on loom (5:10): a technician starts a live call and says what's wrong; offshore listens, looks things up, writes the work order and emails everyone.

a pitch slide: "offshore turns one issue report into coordinated action". three steps: understand (search oem manuals, maximo history, api standards and field context), draft (a work order with asset id, failure mode, parts and safety notes), act (email the planner, oem rep and procurement team and follow the thread to closure). footer: nia is the context layer, agentmail is the action surface — one issue report becomes coordinated action: understand, draft, act.

offshore.

github demo video on loom rednote (小红书) post on rednote

apr 2026

neuro ad-visor @ perplexity ai computer hackathon, with joshua, naisha, jacob

predict how people's brains react to your content before you post it, using meta tribe v2.

first place, most creative computer workflow a gold medal on a red ribbon, with a little crying cat poking its face, paws and feet out of it

the neuro ad-visor demo video, on google drive. the still shows the temporal brain reaction screen: a video of an orange cat on a rug on the left ("wait for it…"), and two red-and-yellow heatmaps of the brain's cortex on the right, lighting up as the video plays — the demo video (1:51, on google drive): upload a clip, watch the brain react, get suggestions.

the brain watching along: as the clip plays, the predicted cortex lights up next to it, every 2 seconds.

the upload screen: "know what converts before you post". a cat video is loaded, the platform is set to instagram, and a big cyan button says predict audience response — drop in a video or picture, pick the platform, press predict.

the overall cortical activation: four views of the brain as red-orange heatmaps, and below them a temporal brain journey chart with lines for fear, anxiety, reward, visual attention, social connection and empathy across the video — the whole cortex from four sides, and nine feelings over time.

an emotional radar, and the perplexity agent explaining what the brain data means.

optimization suggestion cards with toggles: make the first frame instantly legible (image), turn suspense into a story, not just a tease (text), pay off the tension faster (layout), use sound to convert attention into delight (audio), each with its expected impact and target emotion — suggestions to make it land better, each with the feeling it targets.

github demo video read more :)

apr 2026

field: an eeg foundational model @ the tech bros, with iana

field lab's mark: a stippled purple and navy brain made of tiny dots on cream, with the words onto intelligence in serif type below — field-lab.org: onto intelligence.

two eeg recordings of the same cognitive task on a dark background: on the left a dense 64-channel recording (physionet), on the right just 3 orange channels (bnci2014_004) — the same thinking task recorded by two headsets: 64 channels vs 3.

two topographic scalp maps side by side: the 64-channel one full of red and blue blobs and contour lines, the 3-channel one just a smooth red-to-blue split down the middle — what the head looks like with 64 electrodes, and with 3. rich detail vs almost blind.

three scalp maps: the sparse 8-channel input, the 64-channel ground truth, and the dense field scalpfieldnet reconstructs from the sparse input — scalpfieldnet: 8 electrodes in, the full 64-channel field back out.

three scalp maps in a row: the original field, the sparse input with channels masked out, and the reconstructed field — masking channels and filling them back in: original, sparse input, reconstructed.

a heatmap of one processed 4-second eeg window from physionet eegmmidb (subject S010), 64 channels by 1024 samples at 256 hz, z-scored in red and blue, with the first 4 channel traces drawn underneath — one 4-second window after eegforge's preprocessing: 64 channels, 1024 samples, z-scored, ready for training.

a log-scale chart of cumulative channel-hours over 30 physionet recordings, rising to about 44, with dashed lines for the full eegmmidb dataset (about 2,221 channel-hours, projected) and the zuna-scale target of 2,000,000 channel-hours — how much eeg we'd gathered (in channel-hours) vs the full dataset and zuna's 2,000,000 channel-hour scale.

a line chart of nmse on held-out channels vs channel dropout rate (20 to 90%) on eegmmidb: zuna's released 380m weights, our trainer fine-tuned for 400 steps, the mne spherical-spline baseline, and our trainer from scratch (64m, 2.5k steps); the fine-tuned model has the lowest error at every rate and spline is worst except for the from-scratch run — replicating zuna: filling in dropped channels. our fine-tuned model beats spline interpolation at every dropout rate.

visit ↗

mar – aug 2026

agentic data validation pipeline @ apple, generative ai @ berkeley

feb – jun 2026

cellquest @ overworld "hack the world(model)", kindred ventures, with anna

enter the vein as a blood cell.

first place overall a gold medal on a red ribbon, with a little crying cat poking its face, paws and feet out of it

cellquest running again, 10 months later, on overworld's waypoint world model and a rented gpu. every frame is generated live as the drug molecule swims through the tissue; 13 hits clear the abnormal cells (40 s).