Senior Software Engineer·Coinbase

Kyle Mitra

Technical lead for Coinbase's LLM Gateway and lead engineer on the Company Brain, building the AI infrastructure behind the company's agents and developer tools. Previously led engineering for Coinbase One launches, including the Coinbase One Card and the first Member Week.

Duke University · Biomedical Engineering & Computer ScienceNew York

Career Timeline · 2022 – 2026

Joined Coinbase

Joined Coinbase as a new-graduate software engineer on Coinbase One, Coinbase's paid membership. Members pay a monthly fee for zero trading fees, boosted staking and USDC rewards, priority support, and exclusive benefits, so the product combines subscription billing, eligibility, and money movement.

Overview

Selected Impact

Reliability and scale of the LLM Gateway since May 2026, and the membership behind the Coinbase One launches.

LLM Gateway

~86% Fewer Errors

Relative reduction in the gateway's error rate in July 2026 while request volume grew, through cross-provider failover and capacity rebalancing.

LLM Gateway

29 New Models

Onboarded between May and September 2026, for 107 models in the catalog net of retirements, with new frontier models live on release day.

LLM Gateway

7 Model Providers

Hyperscale clouds, first-party model APIs, and specialized inference providers behind one API, serving mainstream frontier models alongside open-weight and open-source models.

LLM Gateway

<2.5 Hours

The fastest end-to-end onboarding of a new frontier model on its release day.

Coinbase One

~1M Paid Members

Coinbase One reached nearly 1 million paid members at the end of 2025, more than 3× in three years, with the Coinbase One Card and Member Week cited among the drivers.

Gateway error-rate figure from July 2026 and model counts as of September 2026. Membership figures from Coinbase's public Q4 2025 shareholder letter.

Projects

Four systems behind AI development at Coinbase, from model access to agent knowledge. Select a component to jump to its case study.

Model calls from internal tools and agents pass through the LLM Gateway; AI Hub, the Strategy Agents, and the Company Brain sit above it. Illustration, not live traffic.

May 2026 – Present · Technical Lead

LLM Gateway

The shared OpenAI-compatible API that Coinbase's internal AI applications and coding agents use to reach models from multiple providers across multiple clouds. Because so much of the company shares it, one provider outage or one ambiguous error reaches everyone at once, so the design goal is graceful degradation and failures that explain themselves.

Ownership

Technical lead and on-call owner since the team took over the service in May 2026. Led a reliability review at handover, wrote the reliability and routing design, and led failover, stalled-stream handling, release gating, model onboarding, access control, caching, and cost attribution.

How It Works

The Product

Clients call a standard OpenAI-style API with one credential and name a model or routing alias; the gateway chooses the provider and deployment, retries, and fails over.
Restricted models use group-based access control.
Every failure carries a stable error class in the body and a header, so a caller can tell a spend limit from a provider throttle or an open circuit breaker. Usage is attributed to the calling team, user, and repository.

Under the Hood

Owned policy on an open-source core. The team owns the policy layer on top of an open-source proxy: authentication, quotas, routing, privacy controls, and audit.
Failover without retry storms. Traffic for each model spreads across providers and regions, unhealthy deployments are taken out of rotation, and failover respects access rules.
Cache affinity. Session affinity keeps repeat calls on one deployment, so provider prompt caching stays effective.
Stalled streams. A watchdog tracks time to first token and idle time between chunks. It shipped observe-only first, then with per-model abort windows, instead of a lower global timeout that would kill long agent runs.
Release safety. The pre-production gate discovers models at runtime, requires one canary per provider family to pass, and separates provider gaps from gateway defects; production rolls out in canary waves against a baseline.

Jul 2026 – Present · Lead Engineer

Company Brain

Coding agents start every task without the institutional memory that lives in pull-request discussions, incidents, and people's heads. The Company Brain distills that memory into a small, human-reviewed Markdown knowledge base for each codebase that keeps itself current and that agents read on demand over MCP.

Ownership

Lead engineer. Authored the technical design and built the system end to end: the distillation engine, the scheduled distillation loop, the admission gate and its calibration set, MCP retrieval, and the evaluation harness.

How It Works

The Product

A repository owner adds a small configuration file with sources and optional domain guidance. An hourly loop then opens pull requests proposing new or updated knowledge pages, each listing every keep or skip verdict with its reason.
The owner reviews and merges like any code change. Nothing enters the Brain without that merge, and every claim links to its source.
During a task, coding agents search, browse, and read the Brain through read-only MCP tools. The engine is versioned, so each team can run its own Brain on the same foundation.

Under the Hood

A human gate enforced by permissions. A multi-stage loop filters, evaluates, writes, and redacts proposed knowledge. Nothing enters without a human merge, and repository permissions enforce that.
Cheapest check first. Inexpensive filters run before any model call, so routine changes never reach a model, and model calls route through the LLM Gateway.
Calibrated in CI. The judge's accuracy is measured against a labeled set and tracked in CI.
Guards against silent failure. Rewrites that drop substantive content are refused, and an empty search raises an error instead of quietly treating every candidate as new.
Many sources, one gate. Pull requests, design docs, and tickets all flow through the same admission and review stages.
Injection hardening. Source text and model output are treated as untrusted at every stage.
Retrieval inside the agent. Loading a whole Brain into context hurt cost and speed, so retrieval runs in the agent's own container against the reviewed copy in its checkout, with no endpoint, token, or tenant-isolation problem.

Apr 2026 · Backend Lead

AI Hub

The company's front door for AI agents. It brings agents from chat, workflow, and search platforms into one catalog with clear owners, and lets employees reuse an account connection across tools.

Ownership

Designed and built the backend: the Agent Catalog service and its ingestion pipeline, delivered from design to production in about two weeks, a Credentials service for delegated account access, and the catalog's listing and detail pages.

How It Works

The Product

One searchable catalog of agents from every platform, filterable by platform and category and tiered as Recommended, Verified, and Community.
A detail page for each agent with its owner, purpose, and source platform, plus a link to open it where it runs.
Connect an external account once through a consent popup, see what access was granted and when it was last used, and disconnect at any time.

Under the Hood

Catalog service. Go, with gRPC and REST over a shared document database, and an agent lifecycle from draft through review, published, deployed, and archived.
Source-adapter ingestion. A scheduled job runs one adapter per platform behind a two-method interface and upserts on source and external ID. If an adapter fails, its archive step is skipped, so an outage cannot wipe that source from the catalog.
Agents versus automations. Workflow platforms host both, so a three-layer classifier checks an explicit production marker, then graph signals such as model nodes and chat triggers, then keywords.
Trustworthy provenance. Records carry a sync-managed flag so sync never overwrites curated agents, and creator identity comes from the signed token, never the request body.
Delegated OAuth. Employees connect external accounts once through a standard consent flow. Tokens stay in a managed secrets store, and every access is authorized and audit-logged.

Apr 2026 · Production Readiness Lead

Strategy Agents

Two internal AI agents let any employee pressure-test an idea: one offers a strategic, big-picture perspective, and the other is a contrarian, first-principles critic. They were built in a week and launched company-wide soon after, so the engineering challenge was making them measurable, testable, and able to scale.

Ownership

Led production readiness for agents another team built: the evaluation and load-testing harness, structured observability and monitors, long-term metrics persistence, per-agent analytics, and a latency design review with a routing proposal.

How It Works

The Product

Message either agent directly, mention it in a thread, or open it in the company's chat interface; replies land in the thread.
Quick questions get a short answer fast. Strategy questions trigger research and a longer, framework-driven reply.
Answers draw on internal documents and recent web sources, with links inline.

Under the Hood

Multi-stage pipeline. Parallel enterprise and web research feeds a perspective stage that applies reasoning frameworks and a separate voice stage, so each stage can be tuned or moved to a cheaper model on its own.
Evaluation harness. A labeled 92-prompt suite covers every route, edge cases, and deliberately ambiguous prompts, and scores routing accuracy and median latency per route. Its pluggable trigger showed that the old and new backends behaved identically.
Load testing that shaped the architecture. The harness's concurrency mode found the prototype workflow engine's limit, which moved the agents to a containerized service with a persistent socket to the messaging platform.
One event per message. Agent, route, latency, status, delivery, and error type drive monitors for silent failures, latency regressions, error rates, and upstream rate limits.
Long-term persistence. A fire-and-forget write in the handler's finally block stores metadata only, never message text, so a database failure can never affect a reply; test traffic is tagged so harness runs never skew analytics.
Latency review. Measured latency per route, sorted eight candidate optimizations into proven and needs-testing, and proposed one shared router that defaults to the fast path.

Part Two2022 – 2026

Coinbase One

Coinbase One is Coinbase's paid membership, with zero trading fees, boosted rewards, a members-only credit card, and more. Paid membership more than tripled in three years to nearly one million at the end of 2025, and passed one million in 2026. Most of its benefits move money, so they have to be exactly right every month.

2025 · Engineering DRI, Onboarding & Membership Integration

Coinbase One Card

Coinbase's first credit card, on the American Express network, with up to 4% back in bitcoin and available only to Coinbase One members. With an issuing bank and a card platform partner behind it, the card and the membership must stay consistent at every step. Coinbase's Q4 2025 shareholder letter named the card as a growth driver.

Ownership

Engineering DRI for card onboarding and the membership integration across an MVP, an invite-only phase, and general availability. Authored the onboarding, integration, eligibility, caching, and card-membership consistency designs; a separate card team owned card processing.

How It Works

The Product

Eligible members see the card across member home, sign-up, a dedicated card tab, and upsells. During the invite-only phase, members joined a waitlist and were invited in daily batches.
A short application with prefilled details, income, terms, and a soft-pull decision with live progress. Members on a trial or monthly plan upgrade before accepting.
After accepting, members get a virtual card immediately and can add it to Apple Pay, and the card becomes the membership's payment method.

Under the Hood

One eligibility rule. A single shared rule in the company's rule-based eligibility service decides who sees the card, and the UI, backend, and benefit pages all use it.
Caching a hot path. Eligibility runs on every app home load, so each check caches with its own time-to-live, an entry expires at its shortest one, and events keep entries fresh.
Event-driven applications. A durable workflow submits the application, partner webhooks become stream events, and the client polls the application record to show each stage live. Approvals schedule offer-expiry and reminder workflows.
Safe activation ordering. Activation is sequenced so that any failure leaves the card and the membership consistent and safe.
Three-state plan gate. One subscription lookup sorts members into non-member, ineligible plan, or eligible plan, instead of evaluating three separate rules.
Aligned lifecycles. Card and membership lifecycles stay aligned with the card partner, so a card never outlives its membership.

Card Announcement ↗Q4 2025 Shareholder Letter ↗

Q4 2025 · Technical Lead, Five Teams

Member Week

Coinbase One's first member-only event: five days, with one new benefit unlocking each weekday morning, spanning trading, deposits, token sales, the card, and staking. The engineering problem was switching benefits owned by five teams on and off on a precise schedule without code changes.

Ownership

Technical lead across five teams: the technical design, schedule, launch checklist, and runbook, daily pre-launch syncs, audience targeting and schedule configuration, and launch-day command.

How It Works

The Product

A Member Week hub with countdowns and a new benefit each weekday morning: zero trading fees, a deposit bonus, boosted token-sale allocations, card referral rewards, and staking perks.
Non-members saw upsells for each benefit and an automatically applied annual-plan discount.
Benefits applied with no action from the member and reverted to normal plan benefits when each window closed.

Under the Hood

Configuration over new code. Benefits ran on existing configuration and eligibility systems, so each could be scheduled, tested, and switched off without changing trade processing.
Time-window rule with fallback. Unlock and lock times come from the config service for each environment, validated for format and ordering, with safe defaults. The bug bash compressed the whole week into one hour by changing configuration alone.
One schedule, two readers. The countdown API and the backend benefit managers read the same timestamps, so the UI can never advertise a benefit the backend won't grant.
Per-benefit gates. Each benefit was gated by audience and region, and the non-member experience could be controlled on its own.
Hybrid targeting. Warehouse cohorts, accurate but delayed, combine with real-time event cohorts that catch new accounts. The largest audiences were pre-targeted ahead of launch.
A daily launch cadence. Each drop followed one runbook: go-live checks, experiments ramped gradually while the benefit stayed time-locked, an automatic 9 AM ET unlock, and first-hour monitoring by named owners.

Q4 2025 Shareholder Letter ↗

2H 2024 · Backend Technical Lead

Premium Concierge

The concierge service gives each Coinbase One Premium member a dedicated support concierge as a single point of contact. It launched publicly with the Premium tier in December 2024.

Ownership

Backend technical lead for a newly formed team. Wrote the backend MVP design and stood up the service's infrastructure; the MVP landed about 1.5 quarters ahead of plan.

How It Works

The Product

A new Premium member is matched with a concierge automatically.
The app shows the concierge's name, photo, and booking link, with a chat that reaches that person.
If the member cancels, access ends; if the service is unavailable, the app falls back to the standard member view.

Under the Hood

A new service from scratch. Load balancing, DNS, service-to-service authentication, document-database clusters, and a dedicated GraphQL subgraph, built first to avoid a later migration.
Event-driven lifecycle. A listener on the membership event hub handles signup and termination and records each change in an assignment event log, without adding calls to the subscription workflows.
Least-loaded assignment. Each new member goes to the concierge with the fewest active members, with round robin breaking ties, to keep load balanced across concierges.
Chat routing. The assignment syncs to the support CRM over signed tokens, so chats reach the right concierge.
Staying in sync. Lifecycle events keep assignments current, and a reconciliation check catches any drift.
Safe rollout. The feature flag controls only what fetches return, and assignment happens at signup, so widening the flag needed no backfill.

Premium Launch ↗

Q2 2024 · Proposer & Tiger-Team Lead

Onchain Membership

Verified Coinbase One is a public proof of membership on Base, issued through the Ethereum Attestation Service, that any application can read. Base gas rebates were the first member benefit built on it.

Ownership

Proposed the initiative, led a tiger team drawn from four teams, and wrote the attestation and gas-rebate designs; directly responsible for the attestation and for verifying members' linked wallets.

How It Works

The Product

A member connects up to three self-custodial wallets and signs a message; each wallet receives a non-transferable attestation.
Onchain applications check for the attestation, and it is revoked if the membership ends.
Base gas spent while subscribed comes back in USDC early the following month, under one monthly cap across all wallets.

Under the Hood

Issuance. An eligibility policy checks live membership, and the attestation platform signs with managed keys and writes to a revocable schema with a single boolean field.
Revocation. A stream consumer handles membership cancel and pause events and revokes the attestation onchain; resubscribing lets the member attest again.
Gas per wallet. A monthly job over the data warehouse measures eligible Base gas from verified wallets, so rebates work with any self-custodial wallet rather than one wallet app.
Per-member cap. The monthly cap applies per member across all of their verified wallets.
Payout. Rebates queue to a handler with duplicate and eligibility safeguards before a payout service sends USDC, and failures are retried from a dead-letter queue.

Coinbase Verifications ↗Gas Rebates ↗

2022 – 2024 · Builder & Owner

Member Rewards

The payout systems behind three Coinbase One benefits: Boosted Staking, Boosted USDC, and USDC back on Advanced Trade fees. They move real money automatically, so correctness and caps matter as much as features.

Ownership

Built the Boosted Staking payouts inside the staking team's service and owned them for two years, led Boosted USDC across three teams, and designed the Advanced Trade Rebate through four product iterations and a US expansion.

How It Works

The Product

Staked assets and USDC earned a boost paid automatically alongside the standard reward, with the USDC boost applied up to a cap at the time.
On Advanced Trade, members pay the normal fee and get USDC back per fill, up to a fee-free allowance.
Order pages and emails show the rebate as estimated while an order fills and finalized when it completes.

Under the Hood

Home and away ownership. The membership team's boost payout code runs inside the staking service, linked to each period's standard payouts, and was owned end to end for two years.
A cap on a moving rate. The standard rewards rate changes over time, so the boost cap was recalculated each period from that period's standard payout.
Rebates per fill. Advanced Trade fees are volume-tiered, so the benefit charges and then refunds. Fills flow through the trading event stream to a consumer that asks the membership service to check the allowance and pay.
Order roll-up. One order can have many fills, so when an order ends, the membership service sums its fill rebates and tracks fill and order states.
Source of truth. Storing rebates in the core trading database would have coupled order writes to the membership service, so the membership service stays authoritative and serves an on-demand query and the order emails.

Education

Education and Early Accomplishments

Duke University, a U.S. Army internship, and a first engineering project recognized at the Intel International Science and Engineering Fair.

Duke UniversityDurham, North Carolina · Class of 2021

Dual Bachelor's Degree

Biomedical Engineering

Computer Science

Certificate

Innovation & Entrepreneurship

Relevant Coursework

Computational GenomicsSequence alignment, genome assembly, and hidden Markov models, implemented in Python

Data Structures & AlgorithmsAlgorithm design and analysis for core data structures

Signals & SystemsLinear systems and signal processing for biomedical applications

Activities

Men's Club BasketballSociety of Hispanic Professional Engineers

U.S. ArmyInternship · May – Aug 2020

Mobile Application Developer · 101st Airborne Division

A summer internship building a mobile app that centralizes information and improves communication across the 101st Airborne Division.

Designed and developed the React Native app with a small team.
Built the front end in React Native and integrated a Firebase back end for secure multi-factor authentication and file hosting.
Contributed to UI and UX design in Sketch and Figma.

Commack High School · Long Island, New York

Storm-Surge Protection for New York City Subways

2016 Intel ISEF Finalist2016 NYSSEF Best in Fair

During Superstorm Sandy, floodwater poured into the subway through street-level ventilation grates. The project redesigned the grate so a sealing plate slides shut beneath it before a storm surge and stays open for airflow the rest of the time. It was modeled with CAD and computational fluid dynamics, then 3D-printed as a scale model.

Day to day, the grate stays open so the tunnels get fresh air.

Illustration, not the original model. Intel ISEF 2016 in Phoenix brought together more than 1,750 students from 75 countries. At the state fair, the project also won Best in Category for Engineering Mechanics.

Science News for Students ↗Smithtown Matters ↗

Personal

About

Outside of work.

New York Native

Born and raised on Long Island, of Filipino and Puerto Rican descent.

New York Sports

A lifelong Yankees, Knicks, and Giants fan.

20+

Countries Visited

An avid traveler.

ReferencesPublic Sources

Coinbase Q4 2025 Shareholder Letter · Feb 2026 · nearly 1M paid Coinbase One subscribers, Member Week, Coinbase One Card
Coinbase Q2 2026 Earnings Call Transcript · Jul 2026 · over one million paid members
American Express Newsroom: Coinbase One Card · Jun 2025
PYMNTS: Coinbase One Premium Launch · Dec 2024
Coinbase Verifications and Coinbase One Gas Rebates · onchain membership benefits
Science News for Students: Teens Invent Way to Keep Floodwaters out of Subways · May 2016 · Intel ISEF