Vector search, embeddings, and the infra underneath.

Consultant Database Platform Engineer at NationwideGreen Bay, WI

I'm James Christensen — a database platform engineer who ships production retrieval on PostgreSQL + pgvector, and writes the AWS CDK that provisions it. Embedding pipelines, hybrid search, HNSW indexes, 400+ managed instances.

RDS instances managed
400+
CDK constructs shipped enterprise-wide
3
P0 incidents
0
See the workDownload résumé
illustrative embedding spacedrag · hover a point
  • pgvector / retrieval
  • Postgres platform
  • AWS CDK / infra
  • delivery / leadership
01 / selected work

Retrieval systems, shipped.

Search at the top of the page to rank these. I tagged each project with keywords, and the score is how many of your terms it shares: a stand-in for vector similarity, not a real index.

B2B SaaS · flagship

clrframe

terms matched
—
clrframe · product walkthroughvideo → transcript → draft SOP

A document hub for companies that also writes the documentation. Record a narrated video walkthrough of a process and clrframe transcribes it, splits it into steps, and drafts an SOP, so the knowledge in someone’s head becomes a document without anyone sitting down to write one. Photos, videos, existing files, and synced OneDrive folders all land in one library you can ask questions of, every answer cites the passage it came from, and a read-only agent mode researches the library to propose new SOPs.

  • Works from Claude and ChatGPT. A remote MCP server lets an assistant search and read the library with just a URL. Its own OAuth 2.1 server grants scoped read or write access, and no tool can delete anything.
  • Hybrid retrieval. Every upload is extracted, chunked, and embedded into PostgreSQL with pgvector. Full-text and vector results are fused with Reciprocal Rank Fusion, and a similarity floor makes Ask decline instead of guessing.
  • Measured answers. Every answer records whether it was grounded, its best match score, and how many sources it cited. A golden-question suite checks retrieval against real Postgres.
  • A map of the library. Documents are clustered by meaning with cosine k-means, a model names each topic, and the layout stays stable between rebuilds.
  • Built to run. About 60 Lambdas across 16 nested CDK stacks in three AWS accounts, SQS ingestion with dead-letter alarms, and 98.9% backend coverage behind a 90% CI gate. A hibernate mode cuts the idle bill from about $105 to $3 a month.
The clrframe library map: 167 documents drawn as linked points in eight colored topic clusters, labeled NIST Security Standards, NASA Composite Materials, BIS Annual Reports, Environmental Waste Management, Neural Network Research, Polymer Rheology, Natural Language Processing, and Operational Checklists.
clrframe · library mapclustered by meaning
PostgreSQLpgvectorAmazon BedrockRAGRRFMCPOAuth 2.1AWS CDKSQSStep Functionsthree.jsAngular
clrframe.com
Open-source demo

pgvector RAG Demo

terms matched
—

A one-command demo that makes embeddings visible: upload a document, watch each chunk embed live into Postgres behind an HNSW index, see the 1024-dimensional vectors projected to 2D, then ask questions with RAG on or off — retrieved chunks light up in vector space. Runs fully local with GPU models via Docker Model Runner.

PostgreSQLpgvectorHNSWEmbeddingsFlaskAngularDocker
GitHub
Enterprise hackathon · led

ClaimLens

terms matched
—

AI research aid for insurance claims. Historical claims are vectorized into Postgres so a claim number — or a plain-language scenario — returns the most similar past claims, replacing hours of manual research and helping newer associates learn from real precedent.

PostgreSQLpgvectorSemantic SearchEmbeddings
Client production site

VanderLeest Trailers

terms matched
—

Production site for a Northeastern Wisconsin trailer dealership, replacing WordPress. Angular SPA with a Cognito-backed admin CMS so the owner runs content and inventory alone, plus a Bedrock sales assistant that tool-calls live DynamoDB inventory and captures leads. Single CDK stack with per-route throttles and edge caching in front of the LLM endpoints.

AngularAmazon BedrockAWS CDKDynamoDBLambdaCognito
GitHub vanderleesttrailers.com
Enterprise platform

Postgres CDK Construct

terms matched
—

The construct that made PostgreSQL a first-class managed offering at a Fortune 100 insurer: RDS PostgreSQL and Aurora provisioned from a single CDK construct with IAM-driven automation for schema and role creation, per-instance CloudWatch alarms wired to BigPanda, and auto-created ServiceNow incidents. Now backing 400+ instances.

AWS CDKRDSAuroraIAMCloudWatchGitHub Actions
02 / experience

Five years, one Fortune 100 platform.

Jan 2025 — Present
Nationwide

Consultant Database Platform Engineer

  • Led the Tiger Team that standardized Postgres and Aurora across the enterprise, shaping long-term RDBMS strategy.
  • Elevated PostgreSQL to a first-class managed service with an AWS CDK construct for RDS/Aurora and IAM-driven schema and role automation.
  • Managed 400+ RDS instances with per-instance CloudWatch alarms wired through BigPanda into auto-created ServiceNow incidents.
  • Designed and led OpenSearch and Neptune CDK constructs, and migrated repos from Concourse to GitHub Actions.
Jan 2023 — Jan 2025
Nationwide

Specialist Software Engineer

  • Led production support to operational excellence through automation — zero P0 incidents across the period.
  • Built a SOAPUI testing pipeline that automated regression for multiple lines of business.
  • Extended into requirements analysis, plus project, release, and environment management as needed.
Feb 2022 — Jan 2023
Nationwide

Senior Software Engineer

  • Took over as Lead of the ODM production support area and moved monitoring and observability onto BigPanda.
  • Cut production support resourcing 50% through pipeline automation in Concourse, Jenkins, Artifactory, and Harness.
  • Voted Lead of the Automation Guild Pilot Hackathon; presented automation results to Nationwide managers.
Jun 2021 — Feb 2022
Nationwide

Software Engineer

  • Shrank the test case repository 40% with automated cleanup scripting, and caught invalid-data cases that had gone undetected for years.
  • Built a GUI that automated prerelease test-case merging, and a Jira↔GitHub bot that moved cards and linked pull requests automatically.
Jun 2020 — Jun 2021
Nationwide

Intern

  • Delivered a full-time associate workload part-time while finishing a CS degree.
03 / how it works

Document in, cited answer out.

The retrieval path I build and operate — every stage lives in Postgres or a CDK-provisioned AWS service.

  1. Ingest

    Uploads land in S3 and fan out through Step Functions — no request blocks on extraction.

    S3 · Step Functions · Lambda

  2. Chunk

    Documents are extracted and split with overlap, keeping citations back to page and position.

    Textract · Python

  3. Embed

    Each chunk becomes a 1024-dimensional vector and is written straight into Postgres.

    Bedrock Titan · pgvector

  4. Index

    HNSW for vectors, tsvector for lexical — one table, two access paths, one transaction.

    HNSW · GIN · SQL

  5. Retrieve

    Vector and full-text results are fused with RRF, then answered by an LLM with sources cited.

    RRF · Bedrock · agents
hybrid retrieval · rrf over pgvector + tsvector
WITH semantic AS (
  SELECT id, RANK() OVER (ORDER BY embedding <=> $1) AS rank
  FROM chunks WHERE tenant_id = $2
  ORDER BY embedding <=> $1 LIMIT 40
),
lexical AS (
  SELECT id, RANK() OVER (ORDER BY ts_rank_cd(tsv, query) DESC) AS rank
  FROM chunks, websearch_to_tsquery('english', $3) query
  WHERE tenant_id = $2 AND tsv @@ query LIMIT 40
)
SELECT c.id, c.content, c.source_page
FROM semantic s FULL JOIN lexical l USING (id)
JOIN chunks c USING (id)
ORDER BY COALESCE(1.0/(60 + s.rank), 0) + COALESCE(1.0/(60 + l.rank), 0) DESC
LIMIT 8;
04 / stack

What I reach for.

Databases & vector search
PostgreSQLpgvectorAmazon RDSAmazon AuroraSQLHybrid search (FTS + RRF)HNSW indexing
AI & retrieval
Amazon BedrockRAG pipelinesEmbeddingsSemantic searchAgent tool usePrompt engineering
Cloud & IaC
AWS CDKLambdaStep FunctionsAPI GatewayCloudFrontCloud securityObservability
Delivery & leadership
CI/CD automationGitHub ActionsPythonAngularNode.jsMentorshipCross-team enablement
University of Iowa
Education
University of Iowa
BS Computer Science · GPA 3.21
Right now in Green Bay, WI
—Fetching current conditions…
GET /api/weather · lambda
James Christensen
Green Bay, WI · replies within a day
05 / contact

Want to talk about any of this?

Roles, projects, or a question about something on this page.

Download résumé