dist-sys · databases · ai
distributed systems · databases · ai

$ whoami

0xKraken

I build distributed data systems and the AI systems that sit on top of them, and I write about the ways they fail quietly.

4+ yrs in industry M.S. CS · databases open to summer '27 internships

booting…
2.5M+ conversations a week indexed for keyword, semantic and hybrid search
92% query accuracy after replacing LLM-written query DSL with schema-bound tool calls (was 70–85%)
98% precision de-identifying PII, at 10 GB of text an hour
27M records a day batch-loaded from a warehouse into Postgres
500M records a month, plus 10M images, through a fraud-detection ETL off Kafka
2,000 concurrent report sessions after a single-user to multi-user rewrite
~/writing5 posts · all runnable
$ ls -l posts/ --sort=n
$
notes on data systems that fail quietly
step 01

Model the failure

Name the invariant that broke and the mechanism that broke it, in terms that transfer to your system.

step 02

Reproduce it

A simulation, a checker or a script you can run. If a number can't be reproduced, it isn't in the post.

step 03

Fix the class, not the case

Fixes that remove the whole category of bug, and the detector that would have caught it on day one.

~/career · git log --graph
  1. HEAD2026 →

    M.S. Computer Science · databases · open to summer '27 SWE internships

  2. 0x032025 → 2026

    Software Engineer II · big tech · LLM platform

    • +Replaced an LLM-to-query-DSL generator with schema-bound tool calls: accuracy 70–85% → 92%, silent zero-hit failures gone, ~2.5K lines of dead code removed
    • +Root-caused a high-severity database incident to one scheduled query; function errors 9–13% → ~0%
    • +PII de-identification service (ML entity detection + rules-based policy engine): 98% precision at 10 GB/hour
    • +Semantic conversation search over 2.5M+ rows a week: k-NN embeddings behind an agent with keyword, semantic and hybrid modes
    • +LLM-as-judge evaluation pipeline calibrated against human labels, benchmarking 20 adapters across multi-rubric tests
    • +Eval dashboard (filters, cohort views, trends) used by 6 product teams to compare models
  3. 0x022024 → 2025

    Software Engineer III · Fortune 500 retail · analytics platform

    • +Re-architected a single-user platform for many users with session isolation and sync: up to 2,000 concurrent report views
    • +Report loads 60–70% faster through caching and gzip of large CSV payloads
    • +Pluggable data-source layer on Go interfaces, with async, non-blocking queries and live status
    • +RAG chatbot that writes SQL with business context, 90% accuracy
  4. 0x012022 → 2024

    Software Engineer II · Fortune 500 retail · logistics data

    • +Low-latency, real-time notification engine for 150K drivers
    • +Batch pipeline loading 27M+ records a day from a warehouse into Postgres for OLTP
    • +Caching and authorization middleware in Go: API responses 30% faster
    • +Fraud-detection ETL: 10M images and 500M records a month from Kafka streams and third-party APIs
    • +Two engineering awards for data-pipeline work
  5. 0x002017 → 2022

    5-year integrated master's · engineering · ML and big-data coursework

PythonGoSQLC++TypeScriptScalaBashPostgreSQLKafkaSparkAirflowgRPC
KubernetesDockerOpenSearch k-NNBigQueryLambdaRDSFastAPIGinDjangoLangGraphPydanticRAGLLM evals