Data & Platform Engineering

Data platforms, built end‑to‑end — federation to pixels.

I think as a Data Engineer

Given a one-line vision, I scope it, choose the architecture, and ship the whole stack — the data layer and the interface people actually use.

LinkedIn Email

My real stack — storage & query engines → one backend → the surfaces people use.

01 — Approach

No spec. No backlog.
Just a vision — and full ownership.

A one-line vision

"Make use of the data." That was the whole brief.

Embed & discover

Worked beside the people closest to the data to find where it actually hurt.

Decide & ship — solo

Architecture, stack, deployment, roadmap: every call, mine.

Give me a direction, not a spec.

I turn ambiguity into infrastructure — and own every decision in between.

02 — Impact

What changed because the work exists.

<0s

To surface any case

Self-serve — no archive digging, no bespoke query.

Requests through a gatekeeper

Engineers answer their own data questions.

Hands-free QA reporting

Results flow to dashboards & Teams on their own.

0mo

Empty repo → daily driver

Solo, from nothing to a tool teams open daily.

03 — The platform

A debugging & analytics platform, from scratch.

Locked-away telemetry → a self-serve tool teams open every day. ~6 months, solo.

Zero-build frontend

Hand-bundled vanilla JS — near-zero maintenance.

Feature "fences"

Every feature removes in one pass.

Graceful degradation

A missing piece never kills the page.

Federation → ClickHouse

Denormalized hot paths for speed.

Cache + warmer

Compute right-sized to refresh cadence.

Adaptive queries

Fan-out reclassifies heavy/light at runtime.

Shipped into one tool

Archive extractionMedia transcodingCross-source search SSE / NDJSON streamingInteractive analyticsAI-assisted insights Self-serve samplingRBAC & audit

04 — Inside the platform

Four engines under one roof.

On-demand extraction, federated search, live analytics, and self-serve sampling — one FastAPI backend behind them all.

Archive extraction

Production .tar / .tar.gz / .zip / .rif pulled from cloud storage and unpacked on demand.

ffmpeg

NumPy

SciPy matplotlib base64 inline

Session & entity search

Resolve a user, org, or device, then stream its sessions — federated across engines.

ClickHouse

Trino NDJSON / SSE adaptive fan-out

Analytics

Interactive dashboards — data rate, distributions, tiered quality, compare mode, AI insights.

ECharts μPlot

pandas

Claude insights

Sampler

Filter, count, and pull every matching archive with one generated command.

Python

gsutil / GCS

aws / S3

Structured — the pipeline

Scheduled ETL → aggregates

Trino federates the sources; a daily, Dockerized batch job lands rollups in ClickHouse for the dashboards.

Unstructured — on demand

Media & signal

Image frames · video · radar / IQ signal · run logs · nested JSON blobs.

05 — Process automation

QA results that report themselves.

A pipeline I built so quality reaches the team with no human in the loop.

When a run finishes, results land in ClickHouse and refresh Superset — then Teams alerts the channel, linking straight to the exact session in the viewer.

06 — Roles as lenses