ÖUÖmer · online

Working on inference optimization tooling.

fusion.py 1/4

1def rrf(ranks, k=60):

2 return sum(1 / (k + r)

3 for r in ranks)

4# fuse ranks, not scores

press run

$ orchestrator status

workers 200 · healthy

sched p50 12ms

throughput ~50K/h ✓

118ms nDCG 0.371cache off

flags.yaml

formula.tex1/7

θ \leftarrow θ - α \nabla L

gradient step

Cortex-01· neural instrumentSN 0x2A·26

CH 01 · IDLElive

drag to rotateturn the knob

Channel

CH 01 / 06♪ pad

Resting state

40 symbols · 20,000 points

Hold

Vol

32.7767° N · 96.7970° WIndex — 01 / 04

Ömer Ulucan

Building systems that think.

AI · Backend · Applied MathematicsDallas, TX · --:--:-- --

01 — Selected work

Infrastructure that handles real complexity.

01 / 042026AI · Search · Python

Neural Search Engine

Hybrid retrieval for technical documentation, fully local. Every query runs dense BGE-large and SPLADE sparse retrieval in Qdrant and fuses the two rankings with weighted Reciprocal Rank Fusion. A LangGraph router and a local Gemma model turn the results into cited answers.

0.371nDCG@10 on BEIR CQADupStack test (BM25: 0.293)

0.761Recall@100, dev-tuned weighted RRF

50msp50 retrieval: both encoders, Qdrant, fusion

Open case study →

02 — Now

Updated Feb 20, 2026 · Dallas --:--:-- --

Processes currently running.

proc/01

Building

Working on inference optimization tooling — making LLMs faster and cheaper to serve at scale.

1 / 3

proc/02

Learning

Category theory and its applications to software architecture.

1 / 3

proc/03

Reading

Designing Data-Intensive Applications by Martin Kleppmann (re-reading)

1 / 3

proc/04

Thinking About

How to build developer tools that don't fight the user's mental model.

1 / 3

$ htop --user omerload 0.00 · tasks 4 running

PIDCOMMANDCPUSTATE 1042 0% R1059 0% R1076 0% R1093 0% R

03 — Writing

5 essays · click to read

Feb 1, 2026 Why I Build On the difference between understanding systems and building them, and why the latter teaches you things the former never will. philosophy · engineering1 min read →

Jan 10, 2026 On Systems Thinking Why the best engineers think in systems, not features — and how to develop that lens. systems · engineering1 min read →

Dec 5, 2025 The Case for Boring Technology Why PostgreSQL, cron jobs, and monoliths are often the right answer — and when they're not. architecture · pragmatism1 min read →

Oct 18, 2025 Applied Math in Production How linear algebra, probability, and optimization theory show up in everyday backend engineering. math · engineering · AI1 min read →

Aug 22, 2025 Building in Public What I've learned from sharing work openly — the benefits, the risks, and the surprising second-order effects. career · building1 min read →

Let's build · train the network

Wire your problem
into my brain.

input · your problem hidden·

\sum σλ \nabla \to \int L θ

output

tap the inputs on the left · watch the activations

\overset{y}{^}

· forward pass

96.4%P(good fit)

decoded plan

01hybrid retrieval with cross-encoder reranking

est. 5–8 weeks · 18 active weights

Ömer Ulucan

Infrastructure that handles real complexity.

Neural Search Engine

Processes currently running.

Building

Learning

Reading

Thinking About

Wire your probleminto my brain.

Wire your problem
into my brain.