EDWARD7780
Ravi Kumar Gupta, known online as EDWARD7780, is an AI evaluation and training specialist and competitive programmer based in Kolkata, India. I build the environments, reward functions and graders that AI agents are trained and scored on. Open to remote contract work.
The short version
- LeetCode Guardian: contest rating 2127, top 1.39% globally, 1,286 problems solved.
- ICPC 2026 regionalist: team ethOS (Techno Main Salt Lake) was selected for both the Amritapuri regional (Coimbatore site) and the Kanpur regional.
- Merged pull request in microsoft/playwright (#43267, 12 October 2026): fix(reporter): merge json specs across projects regardless of cwd.
- Freelance AI evaluation work in 2026 for Scale AI, RustLabs, Handshake AI, AfterQuery, AirDawg Labs.
- Maintainer of Checkora, an open-source chess platform (139 stars, 497 forks, 134 contributors): merged 175 pull requests and reviewed 167, from 68 contributors, with 287 commits of his own. Its Discord server (https://discord.gg/5nATPn7KRF) had 323 members on 17 November 2026.
- 53 merged pull requests on GitHub.
Competitive programming
- LeetCode (EDWARD7780): Guardian, contest rating 2127, top 1.39% globally, 1,286 problems solved.
- CodeChef (edward_7780_1): 5-star, Division 1, max rating 2080, 232 problems solved.
- Codeforces (EDWARD7780_1): specialist, max rating 1551.
- AtCoder (EDWARD_7780): max rating 1405.
- GeeksforGeeks (ravi5i0bf): 503 problems solved.
- GitHub (EDWARD-012): 53 merged pull requests.
In the ICPC India preliminary contest 2026, team ethOS solved 5 problems and stood 839 of 3,883 teams.
Freelance work
Scale AI: RL Environment Developer, RSI Bench, since Oct 2026
- Build reinforcement learning environments for RSI Bench, used to train and evaluate AI agents.
- Design environment setups, reward functions and automated graders so agent behaviour can be scored reliably.
RustLabs: AI Agent Benchmark Contributor, since Aug 2026
- Build command-line benchmark challenges that evaluate autonomous AI agents in Linux and Docker environments.
- Write reference solutions and tests, and verify each submission is reproducible and hard for frontier agents.
Handshake AI: AI Trainer, LLM Training and Evaluation, since Jul 2026
- Create expert-level prompts, reference responses and grading rubrics used to train and evaluate frontier AI models.
- Deliver submissions that meet project review standards.
AfterQuery: AI Data Contributor, since Jul 2026
- Deliver a high volume of submissions for frontier-model training and evaluation datasets.
- Keep quality consistent across submissions by following detailed project guidelines and rubrics.
AirDawg Labs: AI Systems Evaluator, since Feb 2026
- Review and evaluate Terminal-Bench style dockerized tasks, and verify their outputs against the defined requirements.
- Quality assurance, analytical assessments and documentation-based reviews across AI evaluation projects.
- Draft, refine and review written material for AI-assisted content workflows, to project guidelines and editorial standards.
What I take on
- RL environments and reward design. Environment setups, reward functions and automated graders, built so agent behaviour can be scored reliably.
- Agent benchmark tasks. Command-line challenges for autonomous agents in Linux and Docker, with reference solutions and tests, checked for reproducibility.
- LLM training and evaluation data. Expert-level prompts, reference responses and grading rubrics used to train and evaluate frontier models.
- Rubrics and failure analysis. Structured rubrics for grading model output, preference evaluation, and analysis of where an answer fails.
- Code review. Review of model-generated code in Python and C++.
Open source and projects
- Checkora (Open-source chess platform): A chess platform with an AI opponent. The engine is C++ with a Python fallback and searches with minimax and alpha-beta pruning, behind a Django backend.
- microsoft/playwright #43267 (Merged pull request): Fixed Playwright's JSON reporter so a spec's results from every project merge into one entry whatever directory the tests run from, with a regression test.
- Greedy Trap (Browser puzzle game): Seven optimisation problems where the obvious greedy strategy gets it wrong. An exact solver judges each answer with a contest-style verdict, and every solver is tested against brute force.
- RGPT (AI chat application): A chat app on the Gemini API: a Django REST Framework backend behind a React frontend, with Google sign-in and saved chat history.
- LeaveSync-AI (Leave management system): Leave management with separate dashboards for employees, managers and admins, plus AI leave recommendations.
Skills
- AI evaluation and training: LLM evaluation, AI-agent evaluation, RLHF, SFT data creation, Data annotation, Preference evaluation, Structured rubrics, Failure analysis, Prompt engineering, Adversarial prompting, Code review
- Benchmarks and evaluation design: RL environment design, Reward design, Terminal-Bench, Program Bench, MLE-Bench / ML-Bench, SWE benchmarks, Agentic workflows, Open-ended problems, Numerical answer evaluation
- Programming languages: Python, C++, C, Java, JavaScript, SQL, Bash
- Tools and platforms: Git, GitHub, Linux, Docker, Pytest, Jupyter, LaTeX, Vercel, Render
- Web development: Django, Django REST Framework, React, Next.js, REST APIs, OAuth 2.0, HTML/CSS
- CS fundamentals: Data structures and algorithms, Object-oriented design, Operating systems, DBMS, System design
Contact
ravi5258p@gmail.com · GitHub · LinkedIn
B.Tech in Computer Science and Engineering, Techno Main Salt Lake, Kolkata, 2023 to 2027 (expected).