Berat Ercevik
Software engineer: full-stack products and published research
- Now
- Building products
- Published
- ICML 2026 AIWILD · arXiv
- Open to
- Software engineering roles
2
About
I started coding in seventh grade because I liked making things that actually did something. In high school that grew into CS classes, AP CS, and a hackathon where my team took third. Building with other people clicked for me there, and I wanted more of it.
After high school I built an AI Discord chatbot for a real community. I was obsessed with it. People actually used it, asked it things, and leaned on it in the server, and watching that land taught me how much it means when software reaches someone outside your own laptop.
Then I interned at Trustd.ai before college. That was my first real look at professional software. I worked with a mentor, sat in code review, took feedback that made my work sharper, and watched how production systems are designed, tested, and kept alive. A lot of how I think about shipping came from that room.
When I started at UC Santa Cruz, my curiosity pulled hard toward AI and agents. Through Algoverse I researched how people talk to LLMs in the wild and helped build @GrokSet. I also worked on SkillOptimizer for the ICML 2026 AIWILD Workshop, digging into how agents can get better at skills without heavy task supervision. Alongside that I tutored DSA for upper division students, which forced me to explain hard ideas clearly under real time pressure.
In my software engineering course I joined four teammates on Vitae. We lived in Scrum for months. Timing slipped, tickets collided, and we spent long sessions tracing bugs that only appeared when two features met. We brainstormed until messy ideas became something we could ship, leaned on each other through the rough weeks, and formed the kind of trust you get from solving hard problems together. That project still sits close to me.
This year I built multi-agent pipelines that turn evidence into software, working every day with tools like Codex and Claude. The through line is pretty simple: start curious, put something in front of real people, learn how teams ship, study how agents behave, then build the systems I wanted to exist.
3
Publications
How an agent can get measurably better at a skill by decomposing it into subskills, with no task-level supervision to learn from.
A dataset and analysis of 1M+ public Grok conversations: what changes when people talk to an LLM in front of an audience instead of alone.
4
Experience
Stealth Startup
AI Systems Engineer Intern
Built a multi-agent system that creates software from multimodal, source-based evidence: structured handoffs, persistent execution state, inspectable artifacts, and a sandboxed build-and-repair loop that only stopped when acceptance checks passed. I owned the contracts between agents so a failed step was visible, replayable, and not a silent stall.
- Python
- Google ADK
- GCP
- Docker
UCSC
DSA Tutor
I supported 100+ upper-division students per quarter through office hours and project guidance, explaining data structures, algorithmic tradeoffs, and debugging strategies while collaborating with faculty and teaching assistants on consistent evaluation. Those sessions were a weekly pressure test for walking through recursion, graphs, and complexity out loud.
- Data Structures
- Algorithms
Algoverse
LLM Researcher
Built and analyzed @GrokSet (1M+ tweets of public Grok conversations), fine-tuned BERTopic on conversation-level embeddings, and ran concurrent collection/debug workflows that cut API and compute cost ~50%. The same pipeline later became the paper, so the research and the engineering sat in the same week.
- LLMs
- Python
- Hydra
- Tmux
- Runpod
- SQLite
Trustd.ai
SWE Intern
Shipped admin + REST MongoDB workflows for large user-record sets with Zod validation, then hardened Amplify CI/CD and expanded Playwright/Jest coverage (~60%). That was my first production codebase, and it is why I still reach for tests before I call a feature done.
- React
- NextJS
- TypeScript
- MongoDB
- AWS
- Git
- SCRUM
5
Projects
Jan 2026-Jun 2026
Resume building and version-control platform shipped with a 5-person Agile team. Postgres (Docker/Neon), Clerk auth, GitHub Actions + Netlify CI/CD; ~30% faster API responses, ~70% fewer merge conflicts, 99.9% uptime.
- NextJS
- TypeScript
- Jest
- PostgreSQL
- Docker
- CI/CD
- Neon
- Clerk
- SCRUM
Aug 2024-Sep 2024
AI Discord Chatbot (opens in a new tab)
Llama 3 Discord assistant with multi-agent RAG and self-correction (10K+ indexed messages, 50+ community members).
- Python
- discord.py
- Ollama
- AWS
- LangChain
- SQL
6
Skills
AI & agents
Full-stack
Systems & delivery
Foundations
- Python
- LLMs
- Multi-agent systems
- RAG
- LangChain
- Ollama
- Google ADK
- discord.py
- Hydra
- Tmux
- Runpod
- TypeScript
- React
- NextJS
- PostgreSQL
- MongoDB
- SQLite
- SQL
- Express
- Clerk
- Neon
- Zod
- Docker
- GCP
- AWS
- CI/CD
- Jest
- Playwright
- Linux
- Git
- SCRUM
- Data Structures
- Algorithms
7
Let’s connect.
If you are hiring software engineers, I want to hear from you.