I build LLM applications around business context, tools and human review.

Focus: RAG and retrieval, structured outputs, MCP and scoped tools, evaluation and caching.

Stack: Python · TypeScript · Node.js · FastAPI · PostgreSQL · SQLite · Docker · GitHub Actions · Git · Linux.

Evidence tooling: GitHub Profile Audit — read-only checks for public repository and profile-README consistency, with tests and CI.

• Open Source Radar #169 — merged. Added a Vue 3/Vite quickstart with setup and common install, dev, test, type-check, lint and build commands; linked it from the language index and updated the README quickstart count. The local script suite passed 96 tests, all 6 hosted checks passed, and maintainer TanbirRamim merged it at b5dc243 on 2026-10-16; issue #140 closed as completed.
• NanoCoder #1528 — merged. Fixed assigned issue #1518: when LLM summarization fails, compaction now labels the mechanical fallback as degraded. Reviewer addyCooks approved it; all 17 hosted checks passed, and it merged at 87d8159.
• NanoCoder #1760 — merged. Rejects unknown top-level CLI flags and commands with close-match suggestions while leaving run prompt text free-form. All 18 hosted checks passed; maintainer akramcodez merged it on 2026-10-15 at e313d50.
• NanoCoder #1623 — merged. Preserves literal placeholder-looking text in Markdown responses using exact marker matching, with an entity-decoded collision regression. All 17 hosted checks passed; maintainer akramcodez merged it on 2026-10-15 at ddce40f.
• mcp-memory-service #1534 — merged. Implemented assigned issue #1502, preserving remote tombstone ordering so an older remote deletion cannot override a newer unsynced peer save; maintainer doobidoo merged it into main at ca46032.
• mcp-memory-service #1542 — merged. Implemented assigned issue #1538, adding explicit per-hash keep_local / accept_delete resolution for hybrid tombstone conflicts. Maintainer doobidoo approved and merged it at a2e53c6; 17 hosted checks passed and External Markdown links was skipped.
• mcp-memory-service #1550 — merged. Replaced five dashboard inline click handlers with registered event listeners; the linked issue #1549 was closed by the maintainer on 2026-10-15 after the fix merged. The maintainer merged it at f6bec0a; 18 hosted checks passed and External Markdown links was skipped.
• mcp-memory-service #1561 — merged. Implemented assigned issue #1547 with page-specific Content Security Policy for the dashboard, API overview and SSE test page. At final head 3b8b0f1, the web suite passed 229 tests locally; GitHub reports 20 checks passed and 1 external-link check skipped. Maintainer doobidoo merged it at df3af0d on 2026-10-16; issue #1547 closed immediately after.
• mcp-memory-service #1563 — merged. Scoped semantic duplicate detection to the requested store while keeping exact content hashes globally unique, addressing assigned issue #1517. At final head dc3bc89, tests/storage passed 507 tests (15 skipped, 3 xfailed); 17 hosted checks passed and external-link checking was skipped. Maintainer doobidoo merged it at f572cee on 2026-10-16; issue #1517 was closed after merge.
• NanoCoder #1622 — merged. Fixed Bash completion for assigned issue #1593, where a value-taking flag fell through to subcommand suggestions; merged by the maintainer after focused CLI regression coverage.
• Clinical Deep Research #91 — merged. Replaced diagnostic prints with structured logging and regression tests; updated from maintainer feedback before merge.
• Clinical Deep Research #131 — merged. Made the retrieval package mypy-clean, adopted the current Qdrant query API, and added an in-memory Qdrant lifecycle regression for issue #79. The maintainer merged it at 969fc2a; issue #79 remains open.
• Clinical Deep Research #132 — merged. Migrated the Anthropic provider paths to SDK 1.x with model-compatible sampling, refusal/truncation handling, and fake-server regressions. After responding to maintainer glromero's wire-level review, the final head passed 714 tests and 7 hosted checks; it was merged at 5e61fd3.
• Clinical Deep Research #135 — merged. Replaced synthesis-node console output with severity-aware logging and regression tests for routine and degraded paths. At final head 5dfae30, make check passed 714 tests and all 8 hosted checks passed; merged at c5f2707. The advisory type-check job reported 234 mypy errors across 42 files; attribution to this PR was not established.
• mcp-memory-service #1486 — merged. Implemented assigned issue #1484, splitting search-handler stages and adding a complexity-budget regression.

• Clinical Deep Research #136 — open; requested changes addressed, awaiting maintainer re-review. Preserves PubMed publication labels and normalized study design so the evidence gate can use source metadata. Maintainer glromero tested the branch against 81 PubMed records: it caught 23 prior false positives and classified 8 genuine RCTs that the previous path missed. After review, broad Clinical Trial and Observational Study tags remain unknown, with regressions; the explicit Randomized Controlled Trial mapping remains. Rebased onto current main ec912ea; at exact head 90e7307, make check passes (719 tests) and all 7 hosted checks pass. The prior CHANGES_REQUESTED review awaits maintainer re-review; not merged.
• W&B RAI Toolkit #117 — open, ready for review. Repairs the documented enterprise policy pack for issue #116: consolidates both policies into a schema-valid PolicySet, corrects the CLI/schema documentation, and adds offline regression coverage against the bundled directory. At head 127733a, the full suite passed 716 tests locally with 16 skipped; Ruff, formatting, dependency, REUSE and diff checks passed, as did both hosted Socket checks. The upstream CI workflow is action_required with zero jobs, so no hosted test-suite result is available. No maintainer review yet; issue #116 remains open.
• NanoCoder #1787 — open, review required. Addresses issue #1776: defer loading cli-highlight and its parser dependencies until highlighted output is rendered. At head 529a4f1, the patch changeset and isolated lazy-load regression are included; all 17 hosted checks pass, including Unit Tests & Coverage Analysis. Local pnpm run test:all still reports 52 failures, including CLI child-process ink export resolution and macOS temporary-path assumptions; focused tests for the changed paths passed. Code-owner review is pending; not merged.
• NanoCoder #1794 — open, awaiting code-owner review. Fixes assigned issue #1774: make LSP server discovery non-blocking, add a discovery opt-out, and deduplicate verification by executable. At head ded810b, 228 focused tests pass locally and pnpm test:types passes; all 17 hosted checks pass. The full local suite is not green (35 failures, 15 skipped); a clean origin/main reproduced 34 CLI/path failures, while an additional long-path assertion is worktree-sensitive. A non-blocking automated review found a duplicate-verification gap and a test that did not prove deduplication; follow-up commits address those findings. Three requested code-owner reviews are pending; not merged.
• NanoCoder #1758 — open. Isolates agent-evaluation runs from user preferences and unrelated provider/MCP configuration; 29 focused tests pass. Rebased onto upstream main 4ec94e5 at head 15936fa; all 17 hosted checks pass. Human review is still required; not merged.
• NanoCoder #1645 — re-reviewed; code-owner review pending. After my CHANGES_REQUESTED reproduction on old head c977222, author qwist1233-cpu pushed 0311528, preserving a short failure line alongside its context. On that exact head, the focused AVA file passes 33/33, pnpm run test:types passes, and Biome passes on both changed files. I approved the exact head after verifying that the original failure detail survives compaction. GitHub shows only changeset and path-label checks; no hosted test/build result is available. Three requested code-owner reviews remain; not merged.
• smolagents #2921 — open. Review identified repeated truth-testing of stateful chained-comparison results; the author added a regression and updated the fix. Soukyu approved the exact updated head after the focused 45-test class passed; maintainer review remains pending.
• smolagents #2965 — open. Review verified Python unary-plus semantics with focused regression tests; Soukyu approved the exact head after 2 focused tests passed. Maintainer review remains pending; no hosted checks are reported.
• smolagents #2982 — open; reviewer-approved by Soukyu. Addresses issue #2951: reject malformed native tool-call JSON without rejecting valid scalar strings. At head cc88368, all 10 tool_calling tests pass locally; Ruff check and format pass. GitHub reports no checks, maintainer review is still required, and the PR is not merged.
• smolagents #2993 — review submitted. Independently reproduced a nested-comprehension closure mismatch at head 7ea3ee1: CPython 3.14 returns [[2, 2], [2, 2], [2, 2]], while LocalPythonExecutor returns [[0, 0], [1, 1], [2, 2]]. Submitted a CHANGES_REQUESTED review with the regression; no hosted checks or maintainer review are reported, and the PR remains open.
• Chroma #7831 — open, reviewer-approved. Reviewed NumPy scalar handling for embedding lists and ran the focused normalization tests (10 passed). Reviewer rescrv approved exact head 3977a66. Latest reported run (rechecked 2026-10-15): 52 checks pass; 18 Python shards fail at Start Tilt services before tests and the required aggregate gate fails; 10 checks are skipped. The PR is not merged.
• DocAtlas #4 — open, maintainer-approved. Clarifies duplicate-filename replacement in Quick Start; maintainer verified 27 Markdown targets. GitGuardian, container, local-links, and tests checks all pass (rechecked 2026-10-15); not merged.
• LangGraphJS #3032 — open. Fixes issue #3010: scoped message, tool-call and value projections now resume after a human-in-the-loop interrupt. Three regression cases reproduce the stale-subscription behavior and verify each projection receives the resumed batch; 86 tests across four focused files pass locally, along with SDK build, type checking, lint and formatting. All 24 hosted checks pass (rechecked 2026-10-15 14:17 ICT); no maintainer review yet and not merged.
• LangGraphJS #3034 — open. Addresses issue #3033: prevents concurrent shallow Redis checkpoint writes from restoring stale state or hiding pending writes. The checkpoint-redis package's 39 tests pass locally; build, Oxlint, Oxfmt and git diff --check pass. All 24 hosted checks pass (rechecked 2026-10-15 18:27 ICT); no maintainer review yet and not merged.
• LeRobot #4989 — open, review required. Validates v3 episode boundary metadata while projecting required columns once and checking only selected episodes. Regression coverage includes an undershot dataset_to_index and loading a selected episode despite corruption in an unselected episode's metadata. At head e1535ae (code tree unchanged from 1af0be3), 20 focused tests and pre-commit hooks pass locally; both commits da6a8c2 and e1535ae credit Anil Kumar as co-author. Label PR passes; no hosted code-test result is reported after the metadata-only rewrite, and the required code workflows await maintainer approval for this fork. No maintainer review; not merged.
• Clinical Deep Research #140 — open, review required. Adds an optional year-based publication cutoff for retrodiction across API, search plan, and Case File evaluation. PubMed uses publication-date filtering; ClinicalTrials.gov is omitted because it cannot honor the cutoff, and reports disclose full-text/model-training leakage limits. At head 9dcd2fe, 723 tests pass locally; all 8 hosted checks pass, including the offline golden-set evaluation (rechecked 2026-10-16 12:50 ICT). No maintainer review; not merged.
• LlamaIndex Agents #807 — open; reviewed by Soukyu. Addresses issue #803: fail promptly when a retry-policy predicate raises instead of swallowing the error and timing out. The exact-head retry/control-loop tests pass 101/101 on CPython 3.14; all six llama-index-workflows package jobs across Python 3.10–3.15 pass. One Docker/DBOS PostgreSQL replay check fails with System database accessed before DBOS was launched and needs maintainer/CI triage. Soukyu approved exact head 402635a; PR is not merged.
• LiteLLM #46706 — draft/open. Fixes issue #46701: removes Azure/Entra auth-only fields from Azure AI Foundry Anthropic request JSON while preserving output_config. The focused regression file passes 8/8 locally and 86/86 reported hosted checks pass at head 54c3710; no live Azure request was made. The CLA Assistant badge still reports not_signed and its comment offers a recheck link; no maintainer review yet, and the PR is not merged.

The selected section above separates verified merged work from current upstream items; an open PR is not presented as an accepted contribution.

Turns shipment exceptions into sourced next actions an operator can review.

Python · FastAPI · SQLite · MCP. Simulated operations · scoped access · human approval.

The separate Agent Gym prototype checks evidence-grounded tool-use proposals against synthetic cases, including policy citations, tenant scope and forbidden-action attempts. It is offline and does not claim live-model quality. Implementation and limits.

Tests whether a cached answer is still safe after its context or permissions change.

Python · Evaluation · Cache invalidation. Reproducible cases · exact / semantic / dependency-aware.

Brings source retrieval, structured drafting and Word/PDF export into one review flow.

TypeScript · Node.js · Zod · RAG. In development · legal-source validation remains ongoing.

Email · Repositories