Kaan Yigit
Brooklyn

Kaan
Yigit.

currently shipping data systems · open to interesting conversations

I'm an engineer who likes when the engineering is the hard part. I build the systems behind a commodities trading desk, and spend a fair amount of time staring at gas curves and asking whether ML belongs anywhere near them.

Born
Istanbul
Studied
UIUC '25
Lives
Brooklyn
Email ↗GitHub ↗LinkedIn ↗Resume ↗Reading list ↗

01About me

Kaan Yigit
Brooklyn, '26

I'm a software engineer working where markets meet infrastructure. The day job is writing the production code, data pipelines and ML systems behind a commodities trading desk; the side work is whatever I'm curious about that week. Right now that's with the Global Commodities team at Uniper; before that I was a backend engineer at Getir.

Most of what I enjoy lives in the translation layer between two domains that don't usually share a vocabulary.

I grew up in Istanbul and studied computer science at the University of Illinois Urbana-Champaign (cum laude, '25). Outside work I read about market microstructure, energy infrastructure and low-latency systems, and I keep a couple of side projects running to learn things the day job doesn't teach me.

02What I'm building

The work splits into three loops, roughly in this order, every day.

01 · Platform

Keep the analytics platform running, then make it better

I work on the desk's analytics platform that traders and analysts use across multiple offices, and I'm driving its migration to Azure Container Apps with a modern CI/CD pipeline. The data side is Snowflake pipelines that ingest daily feeds from third-party commodity data vendors.

02 · Strategy

Pressure-test seasonal trading ideas before they touch capital

The desk surfaces fundamental and seasonal hypotheses, and I run them through the backtesting DSL we built (see card 01) to see how they actually held up. The win is usually killing an idea cheaply, not finding a winner.

03 · ML & Agents

Anomaly detection and research agents

I built an ML-based anomaly detection system on North American natural gas pipeline flows that flags unusual patterns for desk investigation. I'm also building Azure AI Foundry research agents (LangGraph + MCP) that compile a daily briefing on selected oil-producing regions. Still calibrating what the desk actually trusts.

03Selected work

01

Trading-desk tools at Uniper Global Commodities

Trading Analyst & Developer·NY·2024–present·Python, Snowflake, Azure

I build the production code, data pipelines, and ML systems behind the desk. The day-to-day is the analytics platform traders use across NY, London, and Düsseldorf, and the Snowflake pipelines that feed it from third-party commodity data vendors covering production, storage, and pipeline flows.

On top of that sits the work I find more interesting: an ML anomaly detector on North American natural gas pipeline flows that flags shifts for desk investigation, a typed backtesting DSL the desk uses to test ideas cheaply before they touch capital, and Azure AI Foundry research agents (LangGraph + MCP) compiling daily briefings on key oil-producing regions.

StackPython · Snowflake · Azure
OfficesNY · LDN · DUS
Year2025–present
Internal · production code, not public
02

Real-time ADS-B aircraft tracking and collision warning

Senior project, UIUC·2025·Python, Cesium, ML

I led a four-person team building a real-time aircraft tracking system that ingests live ADS-B feeds, runs collision prediction using Closest Point of Approach math, and renders 3D positions in a browser using Cesium.

The interesting bit was the anomaly detection layer: heuristic and ML-based detectors for GPS spoofing, unrealistic vertical rates, abrupt heading changes, position teleporting. The system surfaces color-coded warnings in real time as new positions stream in.

Team4 (lead)
Year2025
StackCesium
github.com/kaanyigit-repo/adsb-prediction ↗
03

Beneath the Surface, multi-task computer vision

Research, UIUC·2024·PyTorch, MTI-Net

A PyTorch multi-task learning framework on MTI-Net that unifies depth estimation, semantic segmentation, and edge detection for indoor scene understanding. Guided attention mechanisms prioritize segmentation features for depth refinement.

The point of the project was understanding how shared representations help multi-task models, not chasing a benchmark number. The architectural choices ended up mattering more than the loss curve.

Depth MAE0.0323
Seg Acc62.17%
DatasetNYU v2
github.com/kaanyigit-repo/beneath-the-surface ↗
04

Backend at Getir

Software engineer·2023–2024·Subscription & checkout

Earlier I was a backend engineer at Getir, the quick-commerce decacorn later acquired by Uber. I worked on subscription and checkout (two revenue-critical paths) and shipped changes that brought renewal latency from 2s to ~100ms and dropped unidentified error rates from 22% to 3.4%.

I also designed the New Relic dashboards and Slack alerting that helped the team move from reactive to proactive on-call.

Latency2s → 100ms
Errors22% → 3.4%
Years2023–24

More small experiments and course projects on github.com/kaanyigit-repo.

04Stack I reach for

Backend
PythonGoJavaTypeScriptSQLPostgresSnowflakeKafkadbt
DevOps & Cloud
AzureAWSContainer AppsDockerCI/CDNew RelicDatadogGrafana
AI & ML
PyTorchscikit-learnLangGraphMCPAzure AI FoundryRoBERTa
Markets
Natural Gas MarketsEuropean GasLNG Trade FlowsStorage DynamicsEnergy FundamentalsWeather

05Currently thinking about

i.

The information gap between energy infrastructure and a real-time trading desk.

A lot of useful work seems to happen when someone bridges two domains nobody bridges, and then writes the boring software that makes the bridge reliable.

ii.

Compiler-style backtesting frameworks.

A strategy as a small DSL, with correctness checks before any data touches it. Less of a notebook, more of a type system.

iii.

How agentic research workflows change the morning meeting.

LangGraph + MCP turn out to be a very practical way to assemble a daily briefing. The hard part is calibrating what the desk actually trusts.

iv.

What gas storage levels actually tell you that the front-month doesn't.

Storage is one of the more honest signals in the European gas curve. Once you've controlled for weather and LNG sendout, what's left tends to be information. Most days the front-month is just catching up to it.

v.

The right amount of weirdness to keep in a personal website.

I think we're close.

06Say hello

Want to talk?

The fastest way to reach me is email. I'm slower on LinkedIn but I do read it. If you want to talk about commodities, low-latency systems, ML on weird data, or you just want to send me something interesting to read, please do.

Send me a note →