Ryan Walsh
Founding engineer and platform architect. Designed Datavault AI's platform.
Platform Architect · Datavault AI
Ch 01 · The system
Grew the codebase from two repositories to ten; founded eight.
- Founding Engineer
- 8 of the company's 10 repositories founded
- 30+ engineers on the platform
63%
Of the code in Datavault AI's platform repositories, authored. More of the running code than the rest of engineering combined. Excluding test code entirely: 61%.
- Infrastructure (Terraform, Kubernetes, GitOps)
- Data-valuation engine
- Customer platform (the rewrite)
- Core API
Rewriting the customer platform
The Next.js platform Datavault AI's exchange runs on. Sole engineer at the start; 30+ engineers build on that architecture today.
Wrote the founding commit and set the architecture.
- Nov 2025 – present
- datavault.one/exchange
- Next.js
- React
- TypeScript
- TanStack Query
- Tailwind
- NextAuth
- Zod
- Vitest
- Playwright
- OpenTelemetry
- Redis
- Docker
What shipped
A server-side memory leak, fixed for good.
A heap snapshot traced OOM-killed pods to cache timers pinning ~1.5 GB of finished server-rendered requests. The fix: a six-line helper, all 57 call sites routed through it, and a build-failing scan for raw values. Zero remain; memory flat through a 7,500-VU soak.
226 files in the founding commit
~27,000 lines of application code, ~34,000 of tests in 43 files, before a second engineer existed. Sole contributor for 59 days.
The core is still original
368 of 368 lines of the base repository class, 953 of 955 of the rate limiter; its seven feature modules are now twenty-four.
A jQuery/EJS admin template became a TypeScript platform.
TypeScript share of code 0% → 94.1%, strict mode with
no-explicit-anyenforced. Tests: 10 files with 179 cases, 2.5% of code → 431 unit files with 5,617 cases plus 21 end-to-end specs with 130 cases, 37% of code. CI coverage floor 4% / 1% / 0.5% → 25% / 60% / 20%, with lint, type-check and the suite on every pipeline. Thirteen hand-vendored libraries (259,387 lines) → none; the codebase it replaced was still 56% verbatim purchased template.
Putting a price on data
IDA, Datavault AI's data-valuation umbrella: where a dataset goes to be priced, bringing together a custom engine and the IBM watsonx.ai-powered agents.
Founded the engine and wrote most of its agents; founded the comparables service.
- May – Aug 2026
- Python
- Temporal
- DuckDB
- Pydantic
- IBM watsonx.ai
- OpenTelemetry
- Playwright
- PostgreSQL
- pgvector
- FastAPI
What shipped
A multi-signal engine
market comparables, a seven-axis utility score from a published framework and a rebuild-cost floor, blended into one range.
A citation auditor
a run fails if a cited comparable is not in evidence.
Cost governance at the call site
a model call that does not declare a prompt-cache breakpoint is refused, with no exemptions anywhere in the pipeline.
The comparables service
scrapers for six data marketplaces, an industry classifier, a human verification queue, a JWT-secured API.
The problem
Data has no price feed, so one model gives false precision. The engine triangulates independent signals, records which one drove the answer, and cites the listings behind them. Caught and reversed a wrong-signed rule after shipping: high rebuild cost was making assets look more valuable. The floor became a true lower bound: you license data, you don’t rebuild it.
A bank rail for real money
ACH money movement on Fiserv's embedded-finance platform, and the screens on top of it.
Built essentially solo; 96% of the surviving code authored.
- Aug – Sep 2026
- Go
- PostgreSQL
- decimal money arithmetic
- Next.js
- React
- TanStack Query
- Zod
What shipped
A 40,000-line Go package built from an empty directory in four weeks
its money routes bound to a money policy by construction, with a separate 986-test suite covering deposits, withdrawals, transfers, statements, onboarding, identity checks, step-up MFA and reconciliation.
A ledger that declares itself a mirror.
The bank’s ledger is the record and the local database only a claim about it; the reconciler re-drives missed webhooks but never moves money on its own evidence, and reports everything else. A reconciler that repairs what it finds is a reconciler nobody can audit.
Failure as a first-class outcome
duplicates reported as unknown, contradictory transitions surfaced rather than absorbed, one idempotency key per intent.
The platform underneath
6
rounds of in-house load testing
Everything Datavault AI's products run on: Kubernetes, Terraform, GitOps delivery, observability, security that fails closed.
Created the infrastructure repository; wrote 99% of its code by hand.
- Sep 2025 – present
- Terraform
- k3s
- Helm
- Argo CD
- AWS
- Traefik
- Kong
- Kyverno
- Sealed Secrets
- Velero
- Prometheus
- Grafana
- OpenTelemetry
- Bitbucket Pipelines
- Go
- Bash
Scale
Seven live environments moved off hand-managed servers in 49 hours
behind written, reversible procedures. None rolled back.
11 environments, 29 Argo CD applications, all defined in the repository
a new environment is a directory.
Production sized from six rounds of in-house load testing, not estimates
a vCPU-per-user rule derived from the last clean run, with the API tier deliberately capped below what the hardware holds because the database’s connection pool is fixed per pod.
Delivery
552 automated deploys across eight environment lanes in ten weeks
each an immutable image promoted unchanged and reconciled by the cluster, from keyless CI, no long-lived key.
337 console-built resources brought under Terraform to a clean no-op plan
17 modules, 95
.tffiles, dev account.The pre-push test gate halved, 102 seconds to 47
by four-way sharding behind a partition guard proving no test escapes.
256 versioned migrations behind a gate
that refuses one numbered below the branch’s maximum — the tool skips those silently.
Security
Authorization that fails closed.
Rebuilt the core API’s route registration: over 400 registrations, every one policy-bound by construction. It will not compile a route without a policy; CI walks the live router and fails on any gap; a runtime deny-by-default check backs it where the pattern resolves.
A cluster that fails closed both ways
default-deny ingress and egress by chart default, five admission policies, nothing in plaintext, production deploys refused.
Alerting written with the counters
a vendor rejecting credentials while health checks read green; OpenTelemetry through a PII-stripping collector.
A knowledge base for engineers and AI agents
What engineers and AI coding agents read before touching the platform, under a citation discipline stricter than code.
Founded both repositories and wrote the founding commit of each; other engineers migrated their own domains in after.
- Jul 2026 – present
- Markdown in git
- MCP
- semantic code search
- local embeddings
What shipped
Two repositories split by audience
266 documents across 19 domains, plus a restricted infrastructure and security set.
An evidence grammar
every load-bearing statement marked verified, refuted or unverified; refuted claims stay in, corrected.
A two-pass refresh
a mechanical update, then a verification pass that caught a subsystem deleted in code but still documented live.
A three-tier context system for agents
path-scoped rules, semantic document search, semantic code search over a self-hosted index.
Bible Map
The whole Bible as a galaxy of people, places, events and books, with quizzes, a timed test and progress tracking.
Designed and built it, end to end.
- Sep 2026 – present
- React
- TypeScript
- Hono
- Zod
What it is
The whole Bible on one timeline
people, places, events and all 66 books drawn as a galaxy from 4004 BC to AD 57, on two scales that meet at the 400 silent years. Built on the open Theographic Bible Metadata set (CC BY-SA 4.0).
A node panel that quotes the ESV
every person and place opens onto the text.
GM Petit Cafe
A static site on AWS for GM Petit Cafe, a classic French pastry bakery in the Austin, TX area.
Designed and built it, end to end.
- Sep 2026 – present
- gmpetitcafe.com
- Astro
- Tailwind
- AWS S3
- CloudFront
What it is
- Built to
dist/and served from S3 through CloudFront. Menu, about and locations
photographed pastries first, type second.
Ch 09 · Index
- 02Nov 2025 – present
Rewriting the customer platform
The Next.js platform Datavault AI's exchange runs on. Sole engineer at the start; 30+ engineers build on that architecture today.
- 03May – Aug 2026
Putting a price on data
IDA, Datavault AI's data-valuation umbrella: where a dataset goes to be priced, bringing together a custom engine and the IBM watsonx.ai-powered agents.
- 04Aug – Sep 2026
A bank rail for real money
ACH money movement on Fiserv's embedded-finance platform, and the screens on top of it.
- 05Sep 2025 – present
The platform underneath
Everything Datavault AI's products run on: Kubernetes, Terraform, GitOps delivery, observability, security that fails closed.
- 06Jul 2026 – present
A knowledge base for engineers and AI agents
What engineers and AI coding agents read before touching the platform, under a citation discipline stricter than code.
- 07Sep 2026 – present
Bible Map
The whole Bible as a galaxy of people, places, events and books, with quizzes, a timed test and progress tracking.
- 08Sep 2026 – present
GM Petit Cafe
A static site on AWS for GM Petit Cafe, a classic French pastry bakery in the Austin, TX area.
Biography
Platform architect, reporting to the CTO; sole architect of the platform. Joined when the company’s code lived in two repositories; ten today.
Authored 63% of the code in Datavault AI’s platform repositories — more than the rest of engineering combined. Excluding test code entirely: 61%.
Leads the QA lead, the DevOps lead and two senior software team leads. Owned the data-valuation and exchange products until product management was staffed.
Skills
- Languages
Go, TypeScript, SQL, Bash, HCL (Terraform), YAML
Working: Python
- Backend
REST API design (Go, chi), Authorization architecture, Payments and ledger engineering
Working: Temporal workflows
- Frontend
React, Next.js App Router, Frontend architecture, TanStack Query, Vitest, Playwright
- Data
PostgreSQL, Schema migrations
Working: DuckDB, pgvector
- Infrastructure
Terraform, Kubernetes (k3s), Helm, Argo CD, AWS, Docker, Environment migration
- CI/CD and release
Bitbucket Pipelines, GitOps deploys, Release engineering, Merge gates
- Observability
OpenTelemetry, OpenTelemetry Collector, Prometheus, Grafana, Alertmanager
- Security
Application authorization, Vulnerability remediation, Trust-boundary analysis, Network and cluster security, Secrets management
- AI engineering
Evidence discipline in AI-written documentation
Working: LLM system design, Coding-agent context systems
- Leadership
Platform architecture, Founding repositories and patterns, Managing team leads, Handoff and succession docs
Before Datavault AI
Engineering since 2019
- GM Petit Cafe · Web Developer Consultant2025–present
Built and deployed a full-stack e-commerce site for a local bakery: React and Tailwind CSS on S3 with CloudFront and an ACM certificate, Node.js and MySQL on Amazon RDS behind EC2, Route 53 routing and IAM-based access. Customers order for pickup and submit catering requests online.
- ISMP Thailand · Co-Founder, Non-Profit2024–2025
Co-founded a non-profit during a one-year sabbatical, mentoring students through English instruction, design workshops and portfolio reviews with local communities.
- Avetta LLC. · Software Engineer II2022–2024
Led a team of three building a microservice that translates the product into 36 languages. Migrated a Scala Cats Effect microservice to Play Framework 3.0.2. Saved $10,000 on a vendor contract by moving API calls to another vendor inside one financial quarter.
- Avetta LLC.2020–2022
Developed and patented a method that maps ESG requirements to supply-chain vendors; over 300 clients and suppliers used it in a year. Built a folder structure with efficient recursive lookups to organize compliance files for thousands of clients. Automated roughly 90% of a microservice's endpoints on Stoplight.
- Avetta LLC. · Junior Devops Engineer2020
Helped migrate Jenkins onto a Kubernetes cluster, configuring StatefulSet, Service and PersistentVolume objects so Jenkins stayed reachable over DNS, kept its data, and resumed jobs after a pod crash.
- Avetta LLC. · Software Engineer Intern2019–2020
Built a multithreaded job runner in Scala on Docker that automated user migration into the system, including 50,000 Amazon client accounts from spreadsheets, with every parsing and deserialization step unit-tested in JUnit.
- Siemens · Database Developer Intern2018
Designed a language-localization database for chemical analyzers that surfaced 500 translations missing from the interface, and scheduled SQL Server jobs to produce the daily report sent to the translation vendor.
- Calico Brands, Inc. · Information Technology Intern2017
Wrote Excel macros that compiled Electronic Data Interchange reports roughly five times faster than the weekly manual process.
- Full DNA Inc. · Database and Quality Assurance Intern2016
Populated a college-search website with statistics for over 5,000 colleges via SQL.
- University of California, Irvine2020
Bachelor of Science, Computer Science