Award Site Teardowns
中文 Switch language

L.I.S.A.

L.I.S.A. is the contact page that Locomotive, a design agency in Montreal, built for itself. Visitors who want to hire the agency, apply for a job or just say hello do not fill in a form. They talk to a 3D character whose head is an old television set. Locomotive finds ordinary contact forms dull and generic, and since 2023 it has replaced its form with this mock AI assistant. Her lines are written and recorded in advance by the team, and the agency's own cheek and self-mockery are the theme of the site.

The site was made by Locomotive together with the French 3D development studio 60fps, and it won Awwwards Site of the Day on 16 September 2026.

Desktop view with blurred and current dialogue lines, pill options, and a figure with a glowing-eyed TV head
The desktop view after the conversation has started. On the left are two lines of dialogue, the blurred previous line above and the current one, “OMG! You're here! I was just manifesting this. Ask away!”, below it, followed by four white pill buttons and a line with an email address. On the right stands the figure with a silver-grey old TV for a head, two glowing round eyes on its screen

01Concept

L.I.S.A. stands for Locomotive's Interactive Super Assistant. After relaunching its website in 2023, Locomotive published its own case study on Awwwards, with a section that introduces her. The article says that once potential clients love your work, the crucial step is getting them to request a meeting, yet contact forms often come across as dull or generic. The team therefore made L.I.S.A., both as a nod to AI culture and as a way for potential clients to enjoy the agency's wit and personality. The article also says she is not a real AI, that her voice came from Google Cloud's text-to-speech service with copy the team wrote itself, and that she would likely evolve over the coming years. Locomotive's contact page still had no 3D character in September 2025, and by February 2026 it had become the TV-headed L.I.S.A. seen today.

The version Awwwards judged sits on its own subdomain, lisa.locomotive.ca, and its conversation is almost identical to the assistant on the main site's Let's talk page, which is L.I.S.A. herself. The header on the subdomain hides the navigation and keeps only the Locomotive logo and a Let's talk link in the top right. On the subdomain, however, that Let's talk link points to an address that does not exist, so a visitor who clicks it lands on a Page Not Found page.

A visitor first sees the black loading screen that every Locomotive page shares, and then the figure lights up against a pale grey backdrop while a line reading [CLICK] TO START bounces around the TV screen. When the visitor clicks, the site starts playing sound, and L.I.S.A. introduces herself and asks what the visitor wants to do. She offers five options, Start a project, Join the team, Drop a quick word, Discover our culture, and the agency's email address, which is copied to the clipboard on click.

The TV-headed figure on grey with [CLICK] TO START on its screen
The screen before the conversation starts. The figure stands in the middle of a pale grey backdrop, the TV screen shows only a pale blue [CLICK] TO START, and of the six indicator lights under the screen only the leftmost one is lit

The Start a project path breaks a project enquiry form into a series of questions, each preceded by a joke. L.I.S.A. asks in turn for the visitor's name, their company and role, whether they want branding or a website, the budget and the launch date. She then asks for a few lines about the project or an uploaded brief, and finally an email address, before sending all the answers to the agency. The budget options change with the project type, with 25k - 50k as the lowest band for branding and 75k - 125k as the lowest for web, so the agency settles the budget inside the conversation. The step that asks about the project also has a What's a good brief? link, which opens a panel titled The ideal digital brief. A thin line along the bottom of the screen grows as the questions advance and shows how much is left.

Project brief step with a text field, an upload box and a What's a good brief? link
Start a project at the step that asks about the project. On the left is a blurred previous line and then the current question, “Tell me what this project's about, or send over a brief.”, with an underlined text field, an upload box reading Drag & Drop your files or Browse and a What's a good brief? link below it, and a white round arrow button at the bottom right

Of the other three paths, two take only a few steps. Join the team lists the open positions, each linking to an outside job page, and also offers freelance work. Drop a quick word asks only for a message and an email. The third, Discover our culture, is the longest path on the site. It splits into five topics, collaboration, everyday agency life, the annual retreat, personal growth and a winning mentality, and each topic is a monologue by L.I.S.A., the longest running to thirteen lines. Every four or five lines she pauses to ask whether the visitor wants to keep going or go back to the topic list. Listening to the whole retreat topic takes nearly three minutes, and while she talks the TV screen shows photos and video from the team's trips. When that topic ends, one of the options reads Join the team and be on the next trip, which takes the listener straight into the job path.

A retreat story line being typed while the TV screen plays a team trip video
One line from the retreat topic in Discover our culture. On the left the line is still being typed out, about each person getting a secret challenge such as convincing someone to sing a 2000s ballad at karaoke. On the right the TV screen plays a video shot on a team trip, crossed by horizontal scan lines

The conversation has a hundred steps in all, and more than sixty of them belong to the culture monologues, with one line each. Most of the other steps carry five to ten alternative lines that say much the same thing, the greeting has twenty-five, and the site picks one at random each time. A visitor who comes back will most likely hear different jokes. Visitors do not have to hear every line either, because a click on the screen while she talks ends the line at once and shows its text in full. The back button at the bottom left and the blurred previous line both take the conversation back one step, and the button at the bottom right turns the sound off, which turns the conversation into a text-only chat.

02Design

The picture is split into two layers. On the left is a column of text and buttons in black and white only, and on the right is the 3D figure standing in front of a pale grey backdrop, with the two layers never overlapping on desktop. The interface uses only pure black #000000 and pure white #FFFFFF, and the backdrop is a grey that is lighter in the middle and darker toward the edges. Apart from the small red tag that flags a wrong answer, the only things in the picture with texture and colour are the figure and her TV screen. As a result the text on the left reads like a chat transcript, and the visitor's eye settles naturally on the talking face on the right.

All the text is set in Helvetica Now Display Regular with no bold, and the layers come from size and blur. Each line of dialogue is 1.73 times the page's base font size, with a line height of 1.1. The base size steps up with screen width, 15 pixels below 1600 pixels, 17 pixels from 1600 to 1999 pixels, then 2 more pixels per step. So a line of dialogue is 26 pixels on a 1440-pixel laptop and about 29.5 pixels on a 1920-pixel screen. The text on the option buttons is only 0.8 times the base size, about half the size of the dialogue. The previous line stays above the current one at the same font size but shrunk to 80 percent and blurred by 4 pixels, so the visitor can see a line is there without being pulled away by it. Locomotive's own display face, PP Locomotive New, a light serif, appears on this page only in the title of the cookie notice.

The interface parts in the left column are black and white as well. The options are fully rounded pill buttons, white with black text at rest and black with white text when the cursor hovers or the keyboard selects them. The email option has a transparent fill and a very faint outline, which sets it apart from the others. An input field is a single underline with its label shrunk to a small line of text above it. When an answer is invalid, a small red tag with white text appears under the line, such as Fill your name, and this is the only place where the interface parts use red. At the bottom left sits a black round back button, and at the bottom right a vertical black capsule holds the sound on and sound off icons. Along the very bottom of the screen runs a black line 4 pixels thick that grows from left to right with the conversation's progress. After each line is typed, a blinking black block cursor stays at its end, like an old computer terminal waiting for input.

The figure pairs a retro appliance with a human body. Her head is an old cathode-ray TV with a silver-grey plastic case and a chrome rim, with headphones hanging on its side and a small row of blue indicator lights under the screen. A few black cables wind around her neck. Her body has pale skin and an off-white knitted turtleneck whose texture is sharp enough to show the loops of the yarn. The light is soft, and the whole figure looks as if it stands in front of a grey paper backdrop in a photo studio. The most unusual decision in the design is that the TV screen is both her face and the site's display window. Most of the time the screen shows two glowing round eyes, sometimes the Locomotive logo. At certain steps it plays a red-toned video of a face or a video of a team member, and when the visitor rests the cursor on a job, the screen switches to that job's animated GIF.

Four TV screens: glowing eyes, a red-toned face video, a team member video and a job GIF
Four kinds of content on the same TV screen, from left to right: the usual two glowing round eyes, a red-toned video of a face, a video of a team member seen from behind with Design & Code printed on the clothes, and the animated GIF that appears when the cursor rests on a job, with a small window of code laid over it

When the screen is narrower than 1024 pixels, the layout switches to a different arrangement. The figure returns to the middle of the picture and moves up, and the dialogue and options go into a white panel at the bottom whose two top corners are slightly rounded. While L.I.S.A. talks, the panel takes only a small strip at the bottom and the TV head is enlarged to fill the upper half of the screen, and when she finishes, the panel rises to reveal the options. On a phone the dialogue is 22 pixels, and the option buttons flip to white text on black so they stand out on the white panel. The sound button moves to the right edge, and the header's Let's talk becomes a Menu button.

Three phones: the start screen, L.I.S.A. talking over a low panel, and options on a raised panel
Three phone screens. On the left, before the start, the TV head fills most of the screen and shows [CLICK] TO START. In the middle L.I.S.A. is talking, the white panel at the bottom shows only her self-introduction, and glitch stripes appear on the TV screen. On the right she has finished, and the panel has risen to show the options as white text on black buttons

03Motion

The first motion a visitor sees is the loading screen shared across Locomotive's site. On black, it first types out lines that grow one at a time, Digital, Digital-First, Digital-First Design and so on, and then a second group ending in Based in Montreal, Canada. Each letter appears 0.01 seconds after the one before it and flashes through five random symbols before settling into the real letter. Each letter stays for about half a second and then scrambles away the same way, while the Locomotive logo at the bottom disappears letter by letter in random order. When the site is opened again in the same browser tab, this text is skipped and only the logo fades.

Once loading is done, the [CLICK] TO START on the opening screen drifts diagonally across the TV screen at a steady speed, like the idle logo of an old DVD player, and bounces off the edges. After the visitor clicks, the figure slides to the right side of the picture in about a second, fast at first and slow at the end, while the first line starts to type out on the left.

The entrance of each line is the site's most distinctive text motion. A black block moves right from the start of the line and reveals one letter every 0.02 seconds, so a line of seventy-odd letters is typed in about 1.5 seconds. Each revealed letter then flickers through random lowercase letters for 0.25 seconds before it settles, so the block always drags a short tail of changing gibberish behind it. When the line is typed, the block cursor at its end starts blinking and waits until her voice finishes, and only then do the options fade in and drift up slightly over 0.3 seconds after a 0.1-second pause. The text is typed faster than she speaks, so a line is often fully typed during the first half of the voice, and the cursor blinks for several more seconds. When the next line arrives, the line she has just said moves up by its own height, shrinks to 80 percent and blurs, becoming the new previous line. The line before that shrinks to half size, keeps rising and fades out. All of these changes finish within 0.3 seconds, fast at first and slow at the end. When the visitor steps back, the motion plays in reverse, the current line fades out and the previous line comes back into focus.

A line being typed behind a black block cursor under a blurred previous line
A moment in the conversation. At the top left is the blurred previous line, “OMG! You're here! I was just manifesting this. Ask away!”, and below it “We believe the best work happens when people” is being typed, its last few letters still scrambled and followed by the black block. On the right the TV screen shows a red-toned video of a face

The figure keeps following the cursor. When the cursor moves to the other side of the picture, her head turns after it with a lag, about halfway in 0.15 seconds and settled at about 0.9 seconds, fast at first and slow at the end. Her shoulders and upper body turn slightly with it. The camera pans gently with the cursor at the same time, which gives the figure more depth. The cables around her neck do not react to where the cursor rests but to how fast it moves. A quick flick of the cursor sets them swinging, and once the cursor stops they slowly hang straight again. Even when the visitor does nothing, her shoulders rise and fall slowly and her head traces small circles, so she seems to breathe.

Sound and picture are designed together. While L.I.S.A. talks, the row of blue indicator lights under the TV screen lights up more or fewer segments with the loudness of her voice, like a level meter. Once the conversation starts, the site loops a piece of background music about 100 seconds long. As soon as she speaks, the music drops to a fifth of its volume within 0.1 seconds. When she stops, it waits 0.25 seconds and then takes 2 seconds to return to its usual volume, so her lines are always clear.

The TV screen keeps changing on its own as well. Horizontal blue stripes, blocky pixelation, a red vertical stripe and small windows of scrolling code appear on the screen at different rhythms, and at rest each of these glitches shows up only now and then. Each time a new line starts, if that step has no video of its own, the site has a 70 percent chance of throwing the screen into a heavy glitch full of code windows and colour blocks. The heavy glitch lasts until that line ends, and the screen returns to normal half a second later. When the cursor sweeps across the screen, the area it passes breaks up into fine dots, like a CRT picture smeared by a finger, and recovers in under a second.

04Engineering

L.I.S.A. does not understand what visitors say, because the whole conversation is a scripted flow table. The 100 steps are written as JSON straight into an attribute of the page, and after load a Vue 3 app reads them and renders the lines, options and inputs. Each step holds its alternative lines, where to go next, and which options it offers with the step each one leads to. It also lists the inputs to collect with their validation rules (checked with Vuelidate), the position of the progress bar, and the videos to play on the TV at that step. Some branches depend on the visitor's earlier answers, so the budget options, for example, are looked up by project type. A step that has only a next step and no options moves on by itself once she finishes, which is why a monologue is really a dozen or so chained steps. The step and data parameters in the URL let other pages drop a visitor straight into a given step, and the example in a code comment takes a visitor directly into the job path with a job ID.

Every line in the flow table also has a voice, and the diagram below shows one step. Every alternative line has its own recording, named after the line's key. The site fetches and decodes the recording and plays it through an AnalyserNode, while a separate loop that runs only 30 times a second averages the volume of 128 frequency bands into one loudness value. That value goes to two places at once, the shader of the indicator lights and the volume of the background music. The code sets a timer for the length of the recording to know when she has finished, the options appear only after the timer fires and the text has been typed, and a click to skip simply stops the recording. Browsers do not let a page play sound before the visitor does something. The site therefore first tries to play the background music and shows [CLICK] TO START only if that fails, so the visitor's click also unlocks the sound. The choice to turn sound off is stored in the browser, and the next visit goes straight into text-only mode.

Diagram of one conversation step: random line, typed text, MP3 into an AnalyserNode, LEDs and music ducking
Diagram of one step in the conversation's flow table. From the step's 1 to 25 alternative lines the site picks one at random, types out the text letter by letter and plays that line's own MP3. The recording goes through an AnalyserNode that reads the volume 30 times a second, lights the indicator lights under the screen and lowers the background music to 0.1 while she speaks. The options or input fields appear only once the voice has ended and the text is fully typed. A click on the screen stops the voice and finishes the text at once

The 3D figure is rendered with three.js and WebGL. The whole character lives in one 2.8 MB glTF model, the body and several cables are skinned to bones, and the figure has about 21,000 triangles in total. The model contains no pre-made animation at all. The head turns, cable swings and breathing described earlier all come from code that sets bone angles directly every frame. The site smooths the cursor position at three different speeds, and subtracting the slow value from the fast one gives the cursor's speed, which is what swings the cables.

The materials of the figure are organised by name. The names the artists gave materials in Blender (Screen, Lights, Skin, Cloth and so on) map straight to shader classes on the web. Each class swaps a few chunks of code into one shared custom PBR shader, so all of them share the same lighting while adding the screen, the indicator lights or the skin on top. The lighting is baked into a 2048-pixel lightmap and an ambient occlusion map, plus an EXR environment map. The extra textures these shaders need, including the lightmap, the glitch texture, the skin's pores and noise and the fabric normal map, ride on an object inside the model that is never rendered. They download with the model, and the code picks them out by material name. The knit uses a free fabric scan from the BlenderKit asset library.

Four textures from the model: a face photo, a glitch texture, a noise map and the lightmap
Four textures from the model file, on a grey background added for legibility. From left to right: the screen material's base map, an upside-down photo of a face whose fragments the glitch blocks borrow, the horizontally streaked glitch texture, the noise used by the skin material, and the lightmap for the whole character, in which the TV case, the screen and the unwrapped turtleneck can be made out

The TV screen is a shader of its own that combines several layers on one flat surface. The right-side texture is mirrored and fills the whole screen, while the left-side texture is cut into 30 by 30 large pixel blocks that cover the lower half during glitches. The screen's videos stream from the video host Mux through hls.js into video textures, the job GIFs go through another library to become textures, and the bouncing line on the opening screen is drawn on an offscreen 2D canvas. Each glitch effect hangs on its own sine curve and appears when the curve crosses a threshold. A single glitch intensity value lowers all of those thresholds, so the heavy glitch described earlier just pushes that value to its maximum while she talks. When the cursor crosses the screen, the site raycasts to find where it lands on the screen and draws a soft dot into a 64 by 64 half-float texture. That texture is rendered back and forth between two copies and keeps 92 percent of the previous frame each time, or a larger dot keeping only 84 percent on phones, which leaves a trail that fades fast. The screen shader then offsets where it samples by a very fine sine grid weighted by the trail, so the pixels scatter along the cursor's path.

Two screen close-ups: clean glowing eyes and an eye scattered into dots by the cursor trail
Two close-ups of the screen during the same idle loop. On the left, with the cursor off the screen, the two glowing round eyes have clean edges. On the right the cursor has just crossed the left eye, whose lower right is broken into fine dots that trail down along the cursor's path

This approach is light on the GPU but heavy on the network. The renderer fixes the pixel ratio at 1, so the picture looks slightly softer than the page text on high-density screens, there is no post-processing at all, and each frame takes only about thirty draw calls. The black loading screen appears within about 0.1 seconds, and the figure appears between about 4.5 and 5.3 seconds. Before the conversation starts, the site has already downloaded about 5.5 MB, 2.8 MB of it the model, whose textures are plain JPEG and PNG that were never converted to a GPU-compressed format. Once the visitor clicks to start, the screen's videos begin streaming at 1080 pixels square and the background music begins downloading as it plays. Within half a minute of starting, the total reaches about 20.8 MB, about 14.4 MB of it video segments, most of which arrive in the first fifteen seconds or so, and the background music file alone is 4 MB. Phones and tablets download the same files.

The conversation is made of real page text and real buttons, so the keyboard and screen readers mostly work. When a visitor presses Tab, focus moves through the options, the sound button and the back button, and a focused button turns black with white text just as it does on hover. However, the page does not announce new lines to screen readers, the sound and back buttons are read out as the untranslated code strings lisa.toggle-sound and lisa.back, and the page has no top-level heading. With the system's reduced motion setting on, only the transitions of the cookie notice and the brief panel stop, while the typed-out lines, the figure's slide and the screen glitches play as usual. Visitors who cannot hear are not shut out, since every line is shown in full as text.

05Scores and Team

Award Source Date Score
Site of the Day Awwwards 2026-09-16 7.48. Design 7.45 / Usability 7.22 / Creativity 7.89 / Content 7.61
Developer Award Awwwards 2026-09-16 7.52

Of the four categories, Creativity scored highest and Usability lowest. In the developer award, Responsive Design scored highest at 8.6 and Semantics / SEO lowest at 6.8.

Credit Role Record (as of 21 November 2026)
Locomotive An independent design agency in Montreal that makes websites and digital brands, and the team behind the open-source smooth-scroll library Locomotive Scroll, with Montreal given as its location on Awwwards Awwwards works 140, SOTY 1, SOTM 4, SOTD 92, HM 133. FWA 14
60fps A French creative development studio whose Awwwards account gives Nancy as its location, making 3D experiences for the web for brands and agencies since 2013 Awwwards works 37, SOTM 1, SOTD 23, HM 37. FWA 40

L.I.S.A. is Locomotive's own contact page, submitted to Awwwards under both names, and the award page does not say which part each of the two made.

06Archive

Award date
2026-09-16
Awards
Awwwards SOTD 2026-09-16, Awwwards DEV 2026-09-16
Site
https://lisa.locomotive.ca/en
Credits
Locomotive, 60fps
Design
interface
Build
realtime
Awwwards
7.48 / 10. Design 7.45 / Usability 7.22 / Creativity 7.89 / Content 7.61. Developer award 7.52. https://www.awwwards.com/sites/l-i-s-a
Stack
Vue 3, three.js r165 (WebGLRenderer), GSAP (SplitText, ScrambleText), Web Audio API, hls.js and Mux, Vuelidate, FilePond, Barba, modujs
Total assets
About 5.5 MB before the conversation starts, 2.8 MB of it the model. Rises to about 20.8 MB within half a minute of starting, about 14.4 MB of it screen video
Speed
FCP 0.10 to 0.12 s (black loading screen), figure appears at about 4.5 to 5.3 s, CLS 0.06 to 0.07. Continuous scroll —