Skip to content
[ Portfolio / 2026 ]--:--:--Africa/Algiers
Open for work — Q3 2026

Buildingtheinterfacesoftomorrow.

[ 01 — The Manifesto ]

I build tools for people who build tools. My work sits at the seam between systems that need to be fast and interfaces that need to be legible — places where the abstraction either holds or it doesn't, where the difference between a good primitive and a mediocre one compounds across every engineer who touches it.

The interesting problems are rarely about which framework. They're about where the boundaries go: what belongs in the kernel of a system and what belongs at its surface, what to make explicit and what to hide, when to pay for a cache and when accuracy matters more. I spend most of my time on those questions and the long, unhurried conversations they require.

After a decade shipping infrastructure that hundreds of engineers depend on daily, I've come to believe the best developer experience is indistinguishable from good taste — opinion, honesty, and a refusal to paper over the parts that should still feel hard.

Aimen Habis

Aimen Habis · Algiers — Spring 2026

[ 02 — Selected Work ]

Selected work, 2020 — 2026.

Showing 4 / 4

[ 03 — Case Study / Helios ]

Helios — deep dive.

Helios replaces a legacy Prometheus + Grafana stack for a fintech processing 40 billion daily transactions. The legacy stack had p99 dashboard-load latency of 4.8 seconds. We brought it to 1.3s without reducing data fidelity, and the system now costs 60% less to run.

−0%

Dashboard p99 latency

0.0M ev/s

Ingestion throughput

−0%

Infra cost / month

−0%

On-call pages

Problem

The dashboards were the on-call problem.

The legacy observability stack had grown organically across three acquisitions. Metric cardinality crossed 18 million active series. Queries that should have taken milliseconds were timing out mid-incident, and the people on-call were getting paged because the tool they used to diagnose pages was itself failing.

“The fastest way to reduce incidents was to make the diagnostic tool trustworthy again.”

Architecture

A column store, a streaming aggregator, and aggressive pre-materialization.

We replaced the TSDB with ClickHouse fronted by a Rust ingest layer that performs online sketch aggregation (HLL, t-digest) and materializes the forty most-queried roll-ups at write time. Query cost became predictable. The frontend switched to a canvas-based renderer so 50k-series overlays remained interactive at 60fps.

ingest/aggregator.rs
pub async fn aggregate(batch: RecordBatch) -> Result<Materialized> {
    let mut sketch = TDigest::with_compression(200);
    let mut roll_ups: HashMap<DimKey, Counter> =
        HashMap::with_capacity(batch.num_rows());

    for row in batch.iter() {
        sketch.add(row.latency_ms);
        roll_ups
            .entry(row.dimension_key())
            .or_default()
            .observe(row);
    }

    Ok(Materialized { sketch, roll_ups })
}
Code: ingest/aggregator.rs, rust

Results

Trust returned before the rollout completed.

By week three of the partial migration, on-call engineers were self-routing investigations to Helios even for services not yet migrated. That informal shift was the real signal. The explicit metrics followed.

[ 04 — Experience ]

A decade of shipping infrastructure.

[ 05 — Writing & Lab ]

Notes & experiments.

Now Building

A streaming query planner for time-series joins in ClickHouse.

Now Reading

"A Philosophy of Software Design" — John Ousterhout.

[ 06 — Contact ]

Have a project? Let's talk.

hello@aimencodexai.tech

Twitter

@aimenhabis

GitHub

aimen-habis

LinkedIn

aimen-habis

Read.cv

aimenhabis