Backend engineer. I build Java / Spring Boot services and Python AI agents, and I ship the frontend when it needs one.

Most of what I know, I learned by building the things below — 26 public projects, from Kafka microservices to a coding agent that resolves 38 of 50 real GitHub issues.

LinkedIn · LeetCode · Email

• Building agents — a coding-agent harness written from raw LangGraph primitives, then pointed at SWE-bench Verified Mini to find out what the patterns are actually worth.
• Running macos-agent — an agentic RAG system betting that precise retrieval beats raw model size on a narrow domain. Live ↗
• Rewriting my portfolio on Next.js 16 + React Three Fiber.
• Deepening DSA (graphs, DP, backtracking) and system design.

Fifty real Django and Sphinx issues. The agent reads the issue, edits the repo, and the patch counts only if the project's own hidden tests pass. First graded pass was 26/50 (52%).

The agent is mine, not a wrapper — a LangGraph StateGraph with a bash sandbox, built in agents. Each instance runs inside its official SWE-bench image with the testbed conda env on PATH and the network off, so the agent can import the project and run its real tests. Agent-written tests are excluded from the graded diff.

The 12-point jump came from reading all 24 failures rather than retuning prompts:

• The agent was flying blind. Bash ran in a bare python:3.11-slim with no project deps, so import django failed and the agent never ran one test of the repo it was patching. Now it runs in the instance's own image.
• Half the "failures" were the grader. Twelve Sphinx instances died at collection with No module named 'roman' — a 2024 docutils release, not the agent. Three of those patches were byte-identical to gold. The grader now pins docutils and Pygments.
• My own prompt was the bug. "Never edit tests" made the agent protect the bug whenever an existing test asserted it. The rule is now: a test that encodes the reported bug gets updated.
• Also fixed: read truncated the middle of any file over 8k chars, and submitted patches included scratch files.

On the 12 hardest that remained, best-of-3 with a blind judge — three independent agents, each one's test edits reverted and the suite re-run before judging — resolves 7/12, with pass@3 at 8/12. Most of the remaining gap is judge selection, not agent capability.

Not a leaderboard number: the prompt includes the FAIL_TO_PASS test names and the test command, which standard SWE-bench runs withhold, and the hard-12 reruns use the dataset's hints_text. It is a number I can reproduce and explain, which is the point.

Languages  Java · Python · TypeScript · JavaScript · C · SQL

Backend  Spring Boot · Spring Security · JPA / Hibernate · FastAPI · Express · REST

AI  LangGraph · RAG · ChromaDB · Groq · OpenAI-compatible APIs · agent evals & benchmarking

Data & Infra  MySQL · Redis · Kafka · Docker · Git · Linux · Vercel

Frontend  React · Next.js · Tailwind · three.js / R3F

26 public projects — 10 Spring Boot services · 6 AI & agent builds · 7 web apps · 3 systems projects in C.

Code by language, across every repo I've authored:

Python       ██████████████████████████████████████  38.0%
TypeScript   ██████████████████████▌                 22.4%
HTML         ████████████▌                           12.5%
Java         █████████▌                              9.3%
JavaScript   ████████▌                               8.7%
CSS          ██████                                  6.2%
other        ███                                     2.9%   MDX · C · Shell · Dockerfile

Bytes undersell Java — ten Spring services are a lot of design in not many lines, and two Python agents are most of that top bar.

Open to backend and AI-engineering work. The fastest way to reach me is email.