AI agents and automation
Assistants that do real work inside the tools you already use: Telegram, WhatsApp, Discord, email and calendars. Scheduled jobs, approval steps before anything risky, and logs of what ran.
AI, software and hardware, built end to end.
I'm Engin Gedik. I build AI agents, machine-learning and computer-vision systems, web products and connected hardware, from the first sketch to something that runs every day.
- Status
- Taking on freelance projects
- Works with
- Startups, small businesses, research teams
- Research
- Quantum Research Lab, Colorado State University
- Based
- Fort Collins, CO · remote anywhere
- Languages
- English, Turkish
§ 01
What I do
Machine learning
Forecasting, classification and fine-tuned language models, including private LLMs that run on your own hardware. Measured on held-out data, not on the data it was trained on.
Computer vision and imaging
Face and physical-feature analysis, image segmentation, and identity-consistent image generation with custom-trained models.
Geospatial and data
Satellite imagery, Google Earth Engine, and joining messy public datasets into one clean table you can trust, with dashboards on top.
Web products
Complete products: accounts, email, background jobs, hosting, monitoring and backups. Built to keep running after launch day.
IoT, firmware and CAD
ESP32 and FreeRTOS firmware, EV-charging protocols, SOLIDWORKS models, and enclosures that are laser-cut, 3D-printed or fabricated in metal.
§ 02
Selected work
Three case studies, then the rest of the ledger. Most of these can be shown live: ask for a demo of any of them.
Terra, crop forecasts and land cover from satellite data
A two-track ML system built in 24 hours for the CSU Geospatial AI Hackathon, judged with USDA, NASA, Databricks, Esri and AWS.
Corn yield forecasting joins six independent sources into a 26,924-row Parquet across five states and twenty years, with per-forecast-date XGBoost regressors at +46.7% RMSE lift over baseline, calibrated uncertainty cones, and a four-tool Claude agent writing cited plain-English reports.
Land cover and greenhouse gases pairs a SegFormer-B1 model fine-tuned on LoveDA (val mIoU 0.5147) with an IPCC and EPA cited emissions calculator and a "what-if" land-use agent.
State-level forecasts land within as little as 0.6% of the USDA ground truth, with the best counties under 0.3% error, a +46.7% RMSE gain over baseline.
An AI operations team that runs around the clock
A team of specialised AI agents on one Linux server, each reachable from ordinary chat apps: a calendar and reminders assistant, a study assistant, a fitness coach that reads Apple Health data, research agents, and a spoken morning and evening briefing. The agents share a group chat where they mention each other, hand work over, and ask a human for approval with a button press before anything sensitive runs.
Built to be trusted. The chat bots run without admin rights, and long build jobs run on their own gateway so they never slow a conversation down. A watchdog checks every service and bot connection every five minutes. A separate auditor agent reviews the briefings, reminders and scheduled jobs three times a day. Problems escalate from Telegram to WhatsApp to a phone call. Backups are encrypted and run nightly.
On its first run the auditor found that every reminder set from Telegram had been failing silently. It was fixed the same night, and a repair job now catches the same fault within ten minutes.
Reading faces and physique from ordinary photos
Analysis. A pipeline that finds faces in everyday photos, crops them, and runs two frozen ONNX models to measure facial features: proportions, symmetry, skin texture, eye colour, and jaw and cheekbone geometry. It also estimates body shape (leanness, waist-to-hip ratio and tone) when the photo shows it. Feature weights start from published research and are refit with a penalised regression under cross-validation. The weights reload while a session is running, so tuning needs no restart.
Generation. The same feature work drives an image-generation pipeline: a LoRA trained on one person, plus trait LoRAs, inside a two-stage ComfyUI workflow. Stage one makes ControlNet-guided edits to a real photo. Stage two creates new scenes from scratch with the same identity.
It runs unattended through long sessions. If it gets stuck it notices, saves evidence of the screen, restarts itself, and reports what happened.
Every photo gets a per-feature breakdown, so each decision can be traced back to the measurements behind it, and thresholds can be changed mid-session.
Also built · Nos. 04–08
Sizing a quantum bet without the measurement
Largest Certified Cycle, Qupacabrathon 2026The task was to certify a nonlocal game on IQM's Emerald quantum chip, but the platform's job API went down, so no team's calibration run ever executed. We rebuilt the missing number from the chip's own calibration record instead: a density-matrix noise model simulated across all 86 coupled qubit pairs. That showed an 8× spread between the best and worst pairs. We placed a calculated bet, spending $19.99 of a $20 budget, and certified the longest cycle at the event.
TuneSummary
Live, open sourceFree listening stats for Spotify users, built from their own data export. The browser unzips the export and uploads only compact play rows to a FastAPI and SQLite backend. Includes email sign-in, Last.fm live sync with connection-health reminders, genres from MusicBrainz, and a recap story drawn on canvas that exports as images or video.
A private LLM trained on your own code
Fine-tuning, localFine-tuned Qwen3-8B on five years of my own code into a local assistant that writes in my style. A Python pipeline built a weighted dataset, QLoRA (via Unsloth) trained it on one RTX 4070 Ti Super, and the model was quantized to GGUF and served privately through Ollama and Open WebUI. No code leaves the machine.
Source privateAsk for a demo →
IoT and CAD design at Imecar
Industry, deployedNine months of hardware shipped to production. I designed parts and enclosures in SOLIDWORKS, laser-cut from aluminium, and wrote firmware for EV charging stations to the DIN70121 and OCPP standards. I also built a smart meeting-room display: an ESP32 running FreeRTOS, with one core driving the LEDs and the other serving the web interface, backed by a Node.js and PostgreSQL service that kept room status live.
Fly 1 and Fly 2, a simulated fruit fly
Research, in progressFly 1 is a Drosophila in MuJoCo physics with a live 3D viewer that can be started from a browser. Fly 2 rebuilds it with the research done first. It starts from the fly's mapped connectome, a wiring diagram of about 176,000 neurons, and every parameter cites its source and is labelled measured, inferred or assumed. The goal is a fly that learns: to dodge a swat, and to seek out food.
Source privateAsk for a demo →
§ 03
How I work
A short call
Tell me what you are trying to do. If I am not the right person for it, I will say so and point you somewhere better.
A written scope
What gets built, how we will know it works, and when. We agree the cost once the scope is clear, not before.
Small, working steps
You see something that runs early, then regular updates you can try yourself, not a big reveal at the end.
A clean handover
You own the code and the accounts. You get documentation and a runbook, so anyone can maintain it, including me if you want ongoing help.
§ 04
Let's build something together.
A project, a demo of anything above, or a role: write a few lines and I will get back to you by email.