What We Delegate to Independent Developers: a PR Instead of an Interview, Claude Code Welcome

Concrete task buckets waiting for contributors: OpenData adapters, ML experiments, frontend, performance, tests. Our only "interview" is your first pull request. AI-assisted code is welcome — we write with Claude Code every day.

What We Delegate to Independent Developers: a PR Instead of an Interview, Claude Code Welcome

In the previous article we announced that we're opening LEX AI as open source. Now the specifics: what tasks sit in the backlog, how they're packaged, why our only "interview" is a first pull request, and why we love Claude Code.

A PR Instead of an Interview

We don't believe in LeetCode, HackerRank, and three-hour whiteboard interviews. They test the ability to solve problems under stress — not the ability to ship working code into a real codebase.

Our filter is simpler: pick an issue labeled good-first-issue or help-wanted, open a PR, go through review. That is our "interview." Except the output stays in production — and is paid if the task is on the price list.

If the PR lands, we already know:

• You read other people's code and match the project's style
• You write TypeScript without crutches and without any casts
• You test changes locally before pushing
• You self-review before sending
• You discuss calmly in PR comments

That's all we need. After that, we talk contract, rate, scope.

We Write With Claude Code Ourselves. AI-Assisted PRs Are Welcome

We're not against AI-written code. On the contrary — we ship dozens of PRs every week written together with Claude Code. Our CI/CD includes Claude agents that auto-fix failing builds on every push to main. So your workflow with Cursor, Claude Code, Copilot, or Codex is not a problem — it's a plus.

What we check:

• You understand every line you submit — even if an agent generated it
• You tested changes locally (docker compose up, not "the agent said it's fine")
• You don't paste generic React boilerplate that doesn't match the architecture
• You remove dead code and placeholder comments before committing

An LLM assistant is a tool like an IDE. It doesn't make you a worse engineer, and it doesn't make you a better one either — it just speeds up the engineer you already are.

Bucket 1 — OpenData Adapters and ETL

We have 15+ government sources integrated: EDRSR, Verkhovna Rada, NACP, OpenReyestr, OpenSanctions, GLEIF, ICIJ Offshore Leaks, HIBP, NVD, INTERPOL, World Bank. Wanted next:

• European courts: rechtspraak.nl (Netherlands, partially done), justice.cz (Czechia), domstol.se (Sweden), curia.europa.eu (Court of Justice of the EU)
• Regulatory registries: FINMA (Switzerland), BaFin (Germany), AFM (Netherlands), CSSF (Luxembourg)
• LATAM: DNRPA (Argentina), JusBrasil (Brazil), InfoTec (Mexico)
• Sanctions delta-sync: incremental OFAC sync with diffs instead of full download

Typical task — 3 to 5 days:

1. Write the adapter in services/opendata-importers/importers/
2. Add checkpoint + resume logic (base class already exists)
3. Write a test with a fixture
4. Add it to the scheduler config

Stack: Python 3.11 async or Node.js, PostgreSQL COPY, shared base/checkpoint/http_client/ip_pool modules already in place.

Bucket 2 — ML Experiments

The most interesting and most expensive bucket. We're looking for contributors on:

• LoRA fine-tuning of jurisdiction-specific models (civil, criminal, administrative) on 1–10M annotated Q | A pairs
• Custom embeddings — fine-tune BGE-M3 on (legal thesis, relevant decision) pairs from our retrieval log
• Citation verification — a dedicated model that verifies whether a cited article of a code actually contains the claimed text
• Router model — a "which tool to call" classifier based on the query, replacing our current rule-based gateway

Stack: HuggingFace, PyTorch, vLLM, optional Vertex AI / SageMaker. GPU comes from our credit pool with Google Cloud / AWS.

Compensation: fixed + bonus on hitting a metric (e.g., >X% preference rate vs baseline).

Bucket 3 — Frontend and UX

lexwebapp — React 19 + Vite + TailwindCSS + Zustand + TanStack Query. Waiting:

• Evidence panel refactor — search results should render in the right panel, not inside chat (multiple issues open)
• Decision diff viewer — side-by-side comparison of two court decisions with similarity highlighting
• Timeline view — case chronology for a single party (sole proprietor / LLC)
• Law-firm dashboard — multi-user view of the team's cases
• Accessibility audit — WCAG AA for all key pages

Difficulty ranges from a 3-day task (timeline view) to a 2-week project (dashboard).

Bucket 4 — Performance and Infra

• PostgreSQL optimization — our DB is 1.17 TB; some queries take 5–10 s; we need time-based partitioning for the cases table
• pgvector HNSW tuning — 65M vectorized decisions, tuning ef_search vs recall
• Redis cache layer — front-cache for heavy aggregations over case statistics by jurisdiction
• Docker image slimming — some images are 2 GB; multi-stage + distroless needed
• CI/CD speedup — local runner builds the monorepo in 12 min, target is 4 min

Bucket 5 — Tests and Documentation

• Playwright E2E for critical flows: signup → Diia auth → search → export → payment
• Jest coverage for services/ in mcp_backend (currently ~45%, target 75%)
• OpenAPI spec for the HTTP APIs of all three MCP servers
• Architecture diagrams in Mermaid in docs/
• API examples in Python / cURL / JS for developers

These are ideal for a first PR. Low risk, fast review, we're always reachable.

What We Don't Delegate

To avoid confusion:

• Production prompts — live in the private secondlayer-core repo
• Billing business logic — Monobank callback handlers, credit deduction, subscription tier resolution
• Anti-abuse heuristics — rate-limiting strategies, behavioral analysis
• Direct client contact — enterprise law firms, government partners
• Legal decisions in content — what the model answers on sensitive topics (handled with lawyers)

Everything else — fair game.

How to Start

1. Clone github.com/overthelex/secondlayer, run docker compose -f docker-compose.local.yml –env-file .env.local up -d
2. Browse issues labeled good-first-issue, help-wanted, bounty
3. Comment on the issue that you're taking it (to avoid duplication)
4. Open a PR — we review within 48 hours
5. Get paid — UAH via bank or USDT, if the task has a price

For ML, OSINT, or performance tasks — we recommend opening a Discussion first to align on approach. Otherwise there's a risk of doing a PR we'll ask you to redo differently.

FAQ

Q: What if I'm new and have never done a PR to open source? A: There's Bucket 5 (tests and docs). A first PR on a README improvement or a new Playwright test is a great entry point. We'll help with review and advice.

Q: How does payment work? A: Before taking a task, check whether it has the bounty or paid label. If yes, the amount is in the description. Otherwise it's a community contribution without payment, but with a mention in the CHANGELOG and credit in the README.

Q: Can I take a large ML task as my first contribution? A: Better not. Start with a 1–3 day task so we both see how it feels to work with our code. After that — it's all yours.

Q: Will you sign an NDA? A: If the task is in secondlayer-core — yes, a simple mutual NDA. For open-source tasks no NDA is needed.

Open repo: https://github.com/overthelex/secondlayer Contributor issues: https://github.com/overthelex/secondlayer/labels/good-first-issue Discussions: https://github.com/overthelex/secondlayer/discussions Contact: vladimir@legal.org.ua

Write a PR, not a cover letter.