I build local-first evaluation infrastructure for coding agents.

Most AI coding demos answer a soft question: can the agent produce something that looks plausible? My work asks the harder engineering question:

Can an agent solve a real task mined from Git history, under hidden tests, with a trace we can inspect and a run ledger we can verify later?

That is the center of this GitHub profile: small, inspectable systems for agentic AI evaluation, RAG stress testing, reproducibility, context engineering, spec-driven workflow design, and correctness-focused developer tools.

If one of these solves a real problem for you, starring that repository helps other developers find the work. Each one is built to stand alone, with local setup, tests, CI, docs, and a clear inspection path.

flowchart LR
  A["Git history"] --> B["PatchGym<br/>mine real coding-agent tasks"]
  B --> C["Hidden tests<br/>oracle patches<br/>validation command"]
  C --> D["Agent run"]
  D --> E["manifest.json<br/>trace.jsonl<br/>report.json"]
  E --> F["TraceWeave<br/>failure forensics"]
  E --> G["SandboxLedger<br/>tamper-evident ledger"]
  F --> L["ProofDeck<br/>static evidence deck"]
  G --> L
  H["Context Crucible"] --> D
  I["SpecMutate"] --> C
  J["RAGNeedle"] --> K["retrieval stress tests"]
  M["SpecForge<br/>trend evidence to spec workflow"] --> N["profile-grade project blueprint"]
  M --> H

git clone https://github.com/nripankadas07/patchgym
cd patchgym
python -m pip install -e ".[dev]"
python -m pip install git+https://github.com/nripankadas07/traceweave
python -m pip install git+https://github.com/nripankadas07/sandboxledger
python -m pip install git+https://github.com/nripankadas07/proofdeck
patchgym demo --keep-dir /tmp/patchgym-proof
traceweave patchgym /tmp/patchgym-proof/runs/oracle --json
sandboxledger ingest-patchgym /tmp/patchgym-proof-ledger.jsonl /tmp/patchgym-proof/runs/oracle
sandboxledger verify /tmp/patchgym-proof-ledger.jsonl
proofdeck build /tmp/patchgym-proof/runs/oracle --ledger /tmp/patchgym-proof-ledger.jsonl --out /tmp/proofdeck-site
proofdeck verify /tmp/proofdeck-site/bundle.json

That flow produces:

• a real mined coding-agent task;
• hidden-test validation;
• manifest.json with commit ids, patch hashes, artifact hashes, return codes, changed files, and totals;
• trace.jsonl for forensic analysis;
• a verifiable SandboxLedger record for the run;
• a static ProofDeck site with a canonical bundle, audit scorecard, attestation file, and artifact Merkle root.

This is the profile thesis in executable form: agent evaluation should leave evidence, not just screenshots.

I use AI heavily, but I do not want AI-assisted software to be judged by vibes. The systems here are built around harder boundaries:

• hidden tests instead of self-reported success;
• traces instead of opaque agent transcripts;
• manifests instead of loose claims;
• hash ledgers instead of mutable screenshots;
• local-first demos instead of hosted black boxes;
• small parsers and utilities with adversarial tests instead of broad, untestable abstractions.

The result is a portfolio with one technical identity:

local-first infrastructure for evaluating, debugging, and hardening coding agents.

These repositories support the flagship stack without competing with it.

Every active repository is expected to have tests, CI, license metadata, issue templates, a pull request template, security notes, contribution notes, and a clear docs or examples surface.

The flagship repositories have Discussions enabled for design questions, evaluation ideas, benchmark comparisons, and integration notes:

• PatchGym discussions
• ProofDeck discussions
• SpecForge discussions
• TraceWeave discussions

Issues are kept for reproducible bugs, docs gaps, and scoped feature requests. For launch planning, copy, channel strategy, and star-growth operating notes, see the star growth playbook.

Last audited on June 13, 2026 across the live public GitHub profile.

Audit notes:

• Portfolio audit
• Deep audit, June 13 2026
• Deep audit, June 6 2026
• Deep audit, May 28 2026

I use AI for scaffolding, test generation, edge-case brainstorming, and first-pass documentation. The architecture, project boundaries, quality bar, final review, and public positioning are mine.

AI-assisted output has to survive source-checkout setup, local tests, CI, security notes, limitation notes, and manual review before it becomes part of the public portfolio. That is why the profile emphasizes reproducible demos and auditable artifacts instead of fake adoption badges or inflated benchmark claims.

• PatchGym: Local Coding-Agent Benchmarks From Real Git History
• Visible Agent Evaluation: Testing The Loop, Not The Demo
• Safe Local-First AI Tooling: Small Systems With Hard Boundaries

This GitHub profile is intentionally code-first. Career credentials, product leadership context, and publication context live on LinkedIn.

For bugs, design questions, or focused collaboration, open an issue on the relevant repository. For profile-level context, use nripankadas07/nripankadas07.