Follow the
right thread.

Python, SQL and automation engineering.

Reproduce the fault. Implement the fix. Hand over checks your team can run.

From the caller to the right customer An illustrative route carries source items D and C through a child workflow to their matching HTTP recipients. Source itemHTTP recipient DCDC Child workflow
A route illustration, built by Blucca.

Successful run.
Wrong customer?

We reproduced the fault, traced the return boundary, and repaired the route. Inspect the saved synthetic executions and real HTTP records.

Read the native repair & checks

Shipped upstream: our Ivanti Saved Search pagination fix was merged by SYNERGY and released in its community-node package 1.10.0. Run the native checks.

US$650 customer-route audit & repair

One route, a focused fix and five agreed acceptance checks. Scope and payment terms agreed first.

A few useful things.

All work in the archive

At the research bench.

Vesuvius: native AFV evaluation.

Benchmark native papyrus-fiber archives against original manual graphs. The preregistered two-ROI study separates raw and cleaned operating points, architecture controls and DP-specific effects, with complete native family checks.

Read the prospective pipeline study / Current-emitter transfer audit / Run the native ObjectStore handoff.

Hutter: learning to compress.

CPU neural learning and causal priors for lossless compression. Reproduce the completed coded studies, including negative results, and inspect the prepared-stream restoration of the original 1 GB input.

Read the research

CPU runtime tools: author-confirmed ARM64 parity and Native Q8 M1 — FP32 activations with constant Q8 weights; the measured Linux AVX2 EPYC static single-token operator takes 0.146 ms (6.70× vs stock two-Q4 Split8; 1.66× vs dense FP32 with the same dequantized weights). OpenVINO BF16 example inputs: a native dtype-and-value check with a small typed-byte fix.

Have a CPU model/runtime mismatch? US$150 CPU parity diagnosis: two pinned configurations, measured differences and a rerunnable report. Send a minimal repro and deadline.

FHERMA ByteCRT: exact wide-polynomial GPU. Stream 128 independent products through four bounded slots. Native whole-call latency falls from 291.519 to 89.809 ms (69.19%); the official three-fresh-seed result passes 3/3 exact at 89.648935 ms. CPU packing, transfers, GPU work and the complete owned result stay timed. Official B128 result / Pinned streaming source. The legacy two-point result retains 40/40 exact at 20 seeds per point.

Test the data your agent hands off. Run the native agent-output labs: preserve per-call purchase fields, surface SQL fetch errors, freeze JSON tool history, and restore complete LangGraph tool batches. LangGraph PR #9560 author adity982 independently reran the complete provider-free lab on Windows 11—all checks passed. Agno issue #11557 author JiaxPeaceLove reran the full SQL-fetch lab on Windows 11—all checks passed, with normal cleanup and no shim. The independently checked SQL fix by adityav1709 was merged upstream on 2026-11-17. Test LlamaIndex journal recovery across two process restarts, with both parallel answer routes and all four stored values checked. Give each CrewAI flow its own budget using a caller snapshot on the released SDK.

SQLGlot merged the WHERE-only CASE membership guard. Our native review checks NULLs, duplicate rows and protected projection/ON paths.

dbt macro annotations: keep cold and warm parses consistent. A two-file fix preserves fully documented generic tests and catches incorrect arguments through saved-manifest reuse. Released 1.12.6 and current-source checks pass; eight focused tests cover documentation and signature changes. Run the free native review.