I’m Abdulkadir, a final-year computer engineering student heading into AI engineering. Coding agents write most of my code; I write the spec, the checks and the review rules, and I own what ships.
Currently on Erasmus at Poznań University of Technology
Work
An assistant for comparing Turkish participation-banking products. It collects campaign and financing texts from bank websites, extracts the numbers with fine-tuned Turkish transformers, and answers questions through a RAG pipeline that cites its sources.
FastAPI · Next.js · PostgreSQL · Docker · Ollama
An orchestrator for coding agents. It gives each task its own git worktree, runs Codex or Claude, checks the result with your own test command, feeds failures back, has the other model review the diff and keeps a privacy-safe log of everything. Runs on existing subscriptions: no API keys.
2 agents with automatic fallback (Codex → Claude)370 tests passing (in the logged run)
TypeScript · Node.js · React
Coding agents start every session from zero, and on a $20 plan the tokens run out fast. Jarvis is my fix: the agent’s memory lives in plain Markdown files in Git, and hooks load only what matters for the current message, so it remembers past decisions without re-reading everything.
~22% less context per message (same three questions, before vs after)1 command to close a session (was 4–6 tool calls)4 bugs caught by a second model’s review (all fixed before merge)
Python · Claude Code hooks · Markdown + Git
same three questions, before and after the change
Drop in a song, get back vocals, drums, bass and the rest as separate tracks, plus an instrumental. The separation model is Meta’s pretrained Demucs v4; I built the app around it: background jobs, in-browser playback with seeking, ZIP export.
Demucs v4 (htdemucs_ft) · FastAPI · Next.js
Watch it run · Dugout
Dugout is the tool I use to hand coding tasks to AI agents. Every task comes with a command that has to pass; the agent saying “done” is not enough. Below are two real runs from its log.
Outcome: Codex added the feature, docs and tests on the first try, the checks passed, Claude reviewed the diff and approved it, and the task finished passed.
Outcome: Codex hit its limit, Claude wrote the code and said it was finished, the checks failed twice, and the task ended failed. Nothing was merged.
Own git worktree
dugout/list-status-filter-797d
Setup passed
Attempt 1 · first try
codex · gpt-6.1-sol
Agent says
I’ll read the list implementation and CLI contract, add the status filter with tests, then run the requested checks.6 tool calls
command pwd && git · command rg -n 'summarizeLog|TaskStatus|status:|list|filter' · command cat packages/core/src/index.ts && · edit docs/contracts/cli.md · command head -12 packages/core/src/api-schemas.ts · edit packages/cli/src/cli.ts, packages/cli/src/list.ts
Agent says
The core schema already defines the five accepted statuses, so I’m reusing it for validation. The filter runs before the limit, and the tests will check every status, mixed-log ordering, and exact output without a filter.5 tool calls
edit packages/cli/src/cli.test.ts, … · command pnpm exec vitest · edit packages/cli/src/list.test.ts · command node scripts/scan-repo.mjs && · command git diff --check
Agent says
All 31 targeted tests passed, along with lint and typecheck. The full suite is still running and has reported failures outside the changed files. The repository scan also failed because the sandbox denied its attempt to spawn Git …7 tool calls
command git diff -- · command ps -eo pid,ppid,comm · command pnpm exec vitest · command pnpm lint && · command python3 - <<'PY' · command rg -n 'timeout|120_000|60_000|300_000|SIGTERM' · command pnpm lint &&
Agent says
Implemented `dugout list --status` with all five allowed values, exit-3 validation, rejection for delegate/eval, and filtering before `--limit`. Updated the CLI contract and verified exact unfiltered output. - Lint, typecheck, …Agent finished · 631.1k tokens
2m 38s
Checks passed
Isolate 44 workers spawned · ~269ms startup each (spawn + environment, per file) at least ~521ms faster with isolate: false — reuses workers across files instead of one per file scan-repo: clean
Review by claude: approved
The diff meets all four acceptance criteria. Validation reuses `taskStatusSchema`, so the allowed values come from one source. `--status` is added to the forbidden lists for delegate and eval, giving …Task passed
1 attempt · 5 files, +72 lines · 3m 0s
Own git worktree
dugout/scan-repo-2694
Setup passed
Attempt 1 · first try
codex · gpt-6-astra
codex hit its usage limit
exitReason: rate_limit · 3s
Handing over: codex → claude
Attempt 2 · fallback
claude · claude-sonnet-5-5
13 tool calls
read vitest.config.ts · read scripts/build-prototypes.mjs · write scripts/scan-repo.mjs · write scripts/scan-repo.test.mjs · edit scripts/scan-repo.test.mjs · bash ×8
Agent says
I wrote `scripts/scan-repo.mjs` and `scripts/scan-repo.test.mjs`, but none of it has been run. This session denied every Bash command with "requires approval": `pnpm lint`, `pnpm typecheck`, `pnpm test` and `node …Agent finished · 226.1k tokens
57s
Checks failed
Duration 7.48s (tests 79%, import 12%, transform 8%, worker 1%) packages/cli/src/providers/review-prompt.test.ts:40 email test… scan-repo: 1 finding(s)
Attempt 3 · retry with the failure fed back
claude · claude-sonnet-5-5
3 tool calls
bash ×3
Agent says
`pnpm lint`, `pnpm typecheck` and `pnpm test` pass (29 files, 370 tests). `node scripts/scan-repo.mjs` exits 1 with one finding in the repo, so I stopped there. Finding - File and line: …Agent finished · 73.7k tokens
11s
Checks failed
Duration 7.13s (tests 85%, import 12%, transform 2%, worker 1%) packages/cli/src/providers/review-prompt.test.ts:40 email test… scan-repo: 1 finding(s)
Task failed
3 attempts · 2 files, +244 lines · 1m 26s
Replayed from real Dugout event logs. Agent messages are quoted as logged, shortened. Times are from the real run.
About
I study computer engineering at Hitit University (2023–2027) and I’m spending this year on Erasmus at Poznań University of Technology.
I started with deep learning and classic ML projects. Today I focus on AI engineering: LLM applications, retrieval and agents.
Internships, AI engineering roles, or a question about one of the projects: email is the fastest way to reach me.
abdulkadiripekk@gmail.com