✓ agent live in productionv4.0 · 3 regions · rollout 100%
live traffic
4k
transcripts in the eval set
100%
rollout, three regions
94%
eval pass rate
live
engagement in progresshow an engagement runs, illustrated
PAST CLIENTS
Coinbase
Decagon AI
ClassDojo
Scotts Miracle-Gro
Dapper Labs
$
$ cat ./client-portfolio
Previous Clients
Fortune 500 companies, unicorns, major AI labs, leading crypto platforms and social media products.
01 / 10
CLIENT_01FORTUNE 500
scotts-miracle-gro/mylawn
MyLawn, Scotts Miracle-Gro's agentic lawn care app for iOS and Android: a chat assistant that turns an address into a care plan for each lawn, with the products to carry it out.
Chat-first onboarding: the Sierra agent runs the lawn assessment as a conversation, starting from verified parcel data for the address
Health score: 0–100 per lawn, weighting nutrition, plan follow-through and hydration, minus logged issues
Care plan engine: lawn profile, live weather and treatment history become tasks, product picks and safety blocks, re-planned nightly
Commerce: a native Salesforce Commerce Cloud shop, Gemini-personalized product pages and optional AutoPilot shipping
Built for agents: every backend operation runs from a CLI, and a Sim Lab tests the logic against 15 personas
Sierraagent behind the chat
0–100health score, per lawn
agentic UIplans built in chat
stack
React Native
Expo
TanStack
Stripe
Hono
Zod
PostgreSQL
Google Gemini
Vercel AI SDK
Sierra Agent SDK
CLIENT_02AI AGENT PLATFORM
decagon-ai/agent-evals
Playbooks, evals and manipulation detection for Decagon's agent platform, used by multi-billion dollar enterprise clients.
agent playbooks · evals
playbooks
tailored to each enterprise client
evals
to test agent quality
manipulation
improved manipulation detection
Playbooks: improved the agent playbooks Decagon runs for its enterprise clients, the instructions that shape how each AI agent handles a customer conversation
Evals: added testing and evals for their chat systems, so changes to an agent can be measured before they reach customers
Manipulation detection: helped improve and scale the systems that catch users trying to push an agent off-script
CLIENT_03CRYPTO EXCHANGE
coinbase/onchain-summer
Frontend for the inaugural Onchain Summer in 2023, the multi-week festival that promoted Base, Coinbase's Ethereum L2, with daily NFT drops from Coca-Cola, Atari and OpenSea.
Daily minting dApp: a new creator featured each day of the festival, with their NFT collection opened for minting, so there was something new to mint every time people came back
Wallet onboarding: wallet setup in under 60 seconds, so people new to crypto could join a drop without already owning a wallet
Onchain: smart contracts behind each daily drop, plus a custom bridge built on the Optimism SDK to move ETH from Ethereum onto Base
stack
React
Next.js
TypeScript
Thirdweb
RainbowKit
Crossmint
Mint.fun
Arweave
Mirror
Twitter/X API
CLIENT_04EDTECH UNICORN
classdojo/ai-qa-pipeline
An AI layer in the CI/CD pipeline at ClassDojo, the $1B+ edtech platform, that turns failing builds into ready-to-work bug tickets, plus a Vitest migration that cut flaky tests by 70%.
CI to ticket: an LLM reads a failing build's output and the change that caused it, then files a structured ticket with a summary, repro steps and the relevant logs, work an engineer used to do by hand
Vitest migration: moved the test suites to Vitest and made them 60% faster
Flakiness: shared state, timing assumptions and order-dependent tests fixed until flaky failures dropped 70%, so a red build meant a real bug again
CLIENT_05SPORTS NFT PLATFORMS
dapper-labs/design-system
A component system built on Chakra UI and shared across Dapper's sports products: NBA Top Shot, NFL All Day and their successors.
Threaded comments: replies nested two levels deep, optimistic updates, soft deletion that keeps a placeholder when replies remain, and a read-only list preview that opens into a full reply drawer
Recent mentions: up to ten recently tagged people appear the moment a user types @, cached per user in AsyncStorage and handed off to search on the next keystroke
Contact sync and friend discovery: a consent-led flow, a full-list sync with contact names removed from the schema, five suggestion states and batched notifications when a contact joins
Notification navigation: the current stack is saved before a notification opens, so Back returns to where the user was, including deep settings screens and in-progress creation flows
stack
React Native
GraphQL
Node.js
MongoDB
CLIENT_07ACQUIRED BY FLYHOMES
zerodown/internal-tools
Four internal tools shipped for ZeroDown: investment memos, offer tracking, lead search and property management.
IC memo tool: turns the data already in the Flask backend into an Investment Committee memo for each offer, an order of magnitude productivity gain for the PM
Offer process tool: moved status updates from phone calls into a web app, with every state of an offer managed in one place
Lead search dashboard: 10 to 15 filters and search for sales and marketing, backed by Python, GraphQL and SQLAlchemy filters, so lead metrics can be quantified
Property management dashboard: 50+ company-owned properties, with property stats, monthly payments and repair setup
“The team stepped in as full-stack engineers for our internal tools team, quickly took a product ownership role and made an immediate impact. What impressed me most was their ability to understand complex business requirements and translate them into intuitive user experiences.”
Laks SriniCTO, ZeroDown
stack
React
GraphQL
Apollo
Ant Design
CLIENT_08NONPROFIT
apparel-impact-institute/climate-data
The Climate Solutions Portfolio: a platform where apparel brands find climate solutions for their factories and track the cost, savings and carbon impact of each one, taken from discovery through design to build.
Discovery to build: research and design ran as a double diamond across three user groups: brands, solution providers and the Aii team
Solutions search: a searchable repository of climate solutions brands can fund in their factories
Brand analytics: spreadsheet calculations replaced by a dashboard of cost per tCO₂e, annual savings, payback period and a breakdown of brand and facility investment
Ingestion engine: a deterministic Node.js/GraphQL engine transposes multi-tab Excel workbooks into a relational schema with no loss of numeric precision, even as workbooks drifted from the template
Hardening: every failure became a failing unit test until a 50-file corpus ingested cleanly
“They were an absolute pleasure to work with. Their work ethic was top-tier, consistently going above and beyond. They are sharp critical thinkers who approach problems thoughtfully, finding creative and effective solutions, rather than just mindlessly completing assigned work.”
Dan XavierHead of Software, Apparel Impact Institute
stack
Nx
TypeScript
MobX
Mantine
Netlify
Fastify
Pothos
Google Cloud Run
PostgreSQL
Sequelize
Cloud SQL
CircleCI
Cypress
Stytch
Sentry
SendGrid
CLIENT_09ACQUIRED BY COINBASE
spindl/web3-attribution
The frontend codebase and customer-facing analytics dashboard for Spindl's web3 attribution platform, built greenfield from an empty repo.
Customer dashboard: key business metrics and user behavior, tracked and visualized in real time
Forms and CRUD: pages across the app, built on React Hook Form and React Select
Auth: credentials-based sign-in through NextAuth on JWT
stack
TypeScript
Next.js
Turborepo
NextAuth
Recharts
CLIENT_10IDENTITY VERIFICATION
certn/background-checks
The frontend for Certn's internal tools and dashboards and for its customer-facing background check products. Certn is an identity verification and background checks company out of Victoria BC.
Customer Success dashboards: the internal tools Certn's Customer Success team runs day to day to track and manage US criminal background checks for its clients
Customer-facing products: the frontend for the background check products Certn's customers use to order checks and review the results
stack
JavaScript
React
Redux
Ant Design
styled-components
$ cat ./process
How a typical engagement runs
PHASESTARTSHIPPED
01Scopingresearch → interview → scope
02Designwireframe → hi-fi → prototype
03Buildbuild → integrate → eval → deploy
04Handoverdocument → pair → hand over
WHAT YOU RECEIVEone deliverable per phase · yours to keep
01scoping
A written scope document with a quote and timeline
02prototype
A high-fidelity prototype you can test
03build
A deployed production release
04handover
Handover documents and code your team can extend
01Scoping
researchinterviewscope
Workflows mapped on site with the people who do the work, and success defined in numbers.
02Design
wireframehi-fiprototype
Something clickable in front of real users before a line of production code.
03Build
buildintegrateevaldeploy
Full-stack delivery with a working deploy from the first commit.
04Handovershipped · yours to run
documentpairhand over
Docs, patterns and a team that can extend the work on its own.
$ ls ./services
From scoping to handover
01/04 · research · interview · scope
SCOPING
Know what you are building before you build it.
Working with your team to understand the problem, the workflow it lives in and the people it affects, then defining the smallest version worth building and how you will know it worked. You get a clear scope, a plan to deliver it and a candid view of the risks.
YOU LEAVE WITH
A written scope document with a quote and timeline
01
Understand the problem first
The workflow and the people in it come first, so the solution fits how the work is done.
02
A written scope you can act on
The problem, the proposed solution, what it will take to deliver and how success will be measured, in a document your team owns.
03
Measurable from the start
Success is defined in numbers before anything is built, so the result is judged on evidence and agents have a measurable target.
04
Honest about what not to build
If a smaller change, or no software at all, solves the problem, that is the recommendation.
$ ./ask --faq
Most asked questions
pixels&bits · askcost
Product engineering with a specialization in shipping AI products end to end.
Start with the people who will use it, not the model. Map the workflow they actually run, build evals from real transcripts so quality is a score instead of a hunch, then do the unglamorous part: auth, data access, legacy systems.