v.2026.1   ▼ scroll
museum ↗
--:--:-- UTC
zac@stack:~$ ./whoami --long
# loading env · melbourne · production agentic AI · snowflake · RAG · knowledge graphs

Zac Plischka. Data Sci & AI Eng

I build production agentic AI — and the substrate beneath the model: the ontologies, governed provenance, and knowledge graphs that make AI memory trustworthy.

status   shipping production agentic AI stack   snowflake · claude agent sdk · langgraph · vllm · python cert   SnowPro Core based   melbourne, AU cv   download pdf ⤓ email   zacplischka@gmail.com ask   chat with my AI, live ↓
zac@stack:~$ tail -f impact.log
50% faster time-to-insight · agentic AI across analytics pipelines
~95% sole engineer · production web platform: 35+ routes, ~28k lines ts in ~2 months
200GB+ on-prem rdbms → snowflake migration · +40% ml-training data access
8+ self-built headless-agent pipelines compiling comms → an AI-maintained knowledge wiki

About

§ 01 about · cat bio.md melbourne, AU · markdown

Snowflake SnowPro Core certified data scientist building production-grade ML systems, loyalty-program analytics, and agentic AI workflows.

The thesis: the substrate beneath the model. A vector database stores facts — it cannot tell you which version is true. I build the ontologies, governed provenance, and guardrails that make AI memory trustworthy.

At Xephyr, consulting for Wesfarmers Health, I architected and shipped Clarence — a two-tier agent knowledge platform serving 12 client tenants over MCP — built supplier campaign measurement frameworks projected to unlock $5M+ in annual sales lift, and built a text-to-SQL business-decision agent with concept resolution and an LLM-as-a-judge eval harness.

Before that, Data Engineer at Lightbulb Partners — migrated 200GB+ datasets to Snowflake and deployed RAG pipelines that cut manual fault-detection review time by 60%. Alongside all of it: four years as a disability support worker — the NDIS domain depth I bring to client work.

I work across the stack: text-to-SQL on Snowflake, GPU-native enterprise AI POCs on Nvidia H100s with LangGraph and vLLM, Power BI dashboards, FastAPI services, and the eval harnesses that make any of it shippable.

→ exit 0

impactsupplier campaign · projected annual sales lift$5M+
shipped Clarence agent knowledge platform · client tenants12
testsClarence suite · 61 files · all green509
memorytense · point-in-time recall vs vector baseline5/5
throughputsnowflake migration · datasets moved200GB+
latencyfault-detection review time reduction–60%
certSnowflake SnowPro Core · COF-C032026
./portrait.asciilive
zac@stack:~$ git log --oneline career/
  1. 2014 init: b2b sales — WHC machinery (~8 yr, concurrent)
  2. 2017 feat: BSc — monash
  3. 2021 feat: disability support — interchange outer east (4 yr)
  4. 2022 feat: master of data science — monash (concurrent)
  5. 2023 feat: data engineer — lightbulb partners
  6. 2025 feat: data scientist / ai engineer — xephyr
  7. 2025 ship: clarence → production · 12 tenants
  8. 2026 tag: snowpro-core · COF-C03 · verify ↗
  9. HEAD shipping production agentic AI — xephyr
zac@stack:~$ cat skills.toml

[languages]

  • python
  • sql
  • typescript
  • javascript
  • bash

[data_and_ml]

  • RAG
  • knowledge graphs
  • hybrid vector search
  • LLM-as-a-judge evaluation
  • synthetic control · a/b testing
  • customer segmentation (RFM)
  • ontology design

[platforms]

  • snowflake + cortex
  • aws bedrock
  • postgresql + pgvector
  • graphiti + falkordb
  • qdrant
  • terraform · docker compose
  • github actions
  • power bi · fastapi · redis

[agent_infrastructure]

  • mcp servers (http + stdio)
  • claude agent sdk · skills
  • headless claude orchestration
  • litellm → bedrock routing
  • fail-closed auth gateways
  • architecture decision records
  • ports-and-adapters DI

Selected work

§ 02 work · ls -la projects/ total 5 · sorted by impact
grep --stack
★ featured 2025 · xephyr · production

Clarence

Two-tier agent knowledge platform — pgvector + Graphiti/FalkorDB — serving 12 client tenants over MCP. Fail-closed Agent-Key gateway, strict tenant isolation, one-command Terraform deploys.

· · · ·
★ featured 2025 · internal

Querus

Text-to-SQL business-decision agent — Qdrant concept resolution maps vague business terms to canonical entities before it writes SQL, with multi-environment warehouse routing and an LLM-as-a-judge harness scoring queries on semantic equivalence.

· · ·
built 2026 · personal · open source

Tense

Temporal memory for AI agents — an MCP server storing knowledge as a bi-temporal graph on Postgres, answering which version is true now or as of any past date. 100% vs 0% against a fair vector baseline on point-in-time questions.

· ·
built 2025 · community

MyPickle

Pickleball court directory for Victoria, with an interactive map and crowd-sourced ratings. Auth-gated reviews and court submissions.

· ·
realtime 2025

Dinder

Real-time collaborative restaurant decision app. Socket.IO over Redis TTL for ephemeral 30-minute group sessions. Built TDD with Vitest + Playwright.

· · · ·

Process

§ 03 how i ship · ./clarence --trace scroll to stream
~/clarence · production trace
 

Ask me — live REPL

§ 04 ask me anything · ./ask-me live AI chat · grounded in CV + writing · refuses out-of-distribution
live   ./ask-me v0.2 · grounded deepseek-v4-flash · via openrouter
# this is a live chat — type your own question, or tap one below
what's the gnarliest thing you've shipped at xephyr?
Clarence — a two-tier agent knowledge platform serving 12 client tenants over MCP, behind a fail-closed gateway. it doesn't demo flashy, but governed memory is the substrate every agent now runs on. boring infra compounds.
zac@stack:~$

Knowledge graph

§ 05 knowledge graph · ./map --concepts drag to rotate · hover or tap a node
live   17 concepts · 21 links · 4 domains
drag · click · ⇧click to trace

Writing, now & open source

§ 06 writing · now · open source

// writing

The 12-Month Clocka data scientist's reflection on AI's exponential rise · nov 2025
Overnight Obsidian clone two-agent harness demo · 85/232 tests passing in one ~6h agentic run · dec 2025
topics in the queue: · agent knowledge bases that don't rot · what concept resolution buys text-to-SQL · LLM-as-a-judge eval harnesses
→ read the article

// now

Architecting agent knowledge bases and retail media analytics at Xephyr for Wesfarmers Health (Priceline).
Productionising Claude Skills + MCP integrations — now standard across team workflows.
Exploring on-prem GPU-native enterprise AI on Nvidia H100s with LangGraph and vLLM.
Reading: agentic eval harnesses, KG retrieval, Snowflake Cortex Analyst patterns.

// open source & profiles

github.com/zacplischkaprofile · all repos
custodian-mcpgoverned agent-memory MCP server · fail-closed auth · tenant isolation
gavelLLM-as-a-judge eval harness for text-to-SQL · public Chinook dataset
tensetemporal memory for AI agents · bi-temporal graph on Postgres · MCP
/museumwalk the whole site in 3d · experiment
linkedin.com/in/zac-plischkaprofile · writing · history
zacplischka@gmail.combest way to reach me
+61 434 831 935Melbourne, AU · UTC+10/+11
→ github.com/zacplischka