Jersey, Channel Islands October 2026

James Balderson

AI native: my career started with Claude in the loop. No degree. Right now: a data studio, a family of local models, and an AI operator that opens pull requests.

  1. 01Overlane1,058 commits 4,068 tests passing
  2. 02yxlm4 fine-tuned models 22% vs a 100% bar
  3. 0342An AI operator PRs #164 and #165
  4. 04AI at workIntern to certified Claude partner

Overlane

The AI-native alternative to the legacy data-prep giants.

Describe the result you want in a sentence. Overlane builds the pipeline (joins, filters, summaries), runs it on your machine, and shows you the data at every step.

It is a drag-and-drop workflow builder in the spirit of the classic enterprise data-prep tools, written with Claude Code. React on top; transforms run off the main thread in a Web Worker, with DuckDB compiled to WebAssembly so 100k to 200k rows are handled client-side. Your data stays on your computer: when the AI helps, it sees the column names and a few sample rows.

There is a desktop build for Mac and Windows, and server connectors for the places real data lives.

Open the public demo overlane.ai

The public demo runs with the AI switched off, by design.

First commit
30 Jun 2026
Commits to 12 Oct
1,058
Co-authored with Claude
686 of 1,058
Merged pull requests
136
Lines of TS, TSX and Python
~116k excl. tests
Test files
315
Vitest tests passing (PR #165)
4,068 / 4,068
Tools on the palette
45+
Connectors
PostgresSQL ServerSnowflakePower BIMicrosoft Graph

yxlm

Small local models that write enterprise data workflows.

Plain English in, a runnable workflow for the market-leading data-prep platform out, from a model small enough to run locally on a Mac.

yxlm is a private family of open-weights models, LoRA fine-tunes of Qwen3.5, each released as an adapter plus a 4-bit GGUF that runs under llama.cpp. Each release trained on more data; from 0.2 onwards that includes the model's own verified attempts.

The training run, release by release

  1. 0.1 4B

    23 Oct

    Pilot on Qwen3.5-4B. 2,258 training rows, 71 steps, about half an hour on one H100.

  2. 0.1 9B

    23 Oct

    First full 9B fine-tune. 9,034 rows (26.6M tokens), 283 steps, about two hours.

  3. 0.2 9B

    27 Oct

    Output-column naming, repairs, and its own verified attempts. 13,506 rows (40.9M tokens), 423 steps.

  4. 0.3 9B

    27 Oct

    A second round of its own verified attempts. 17,727 rows (52.9M tokens), 554 steps, about four hours.

Benchmark: business-style requests (L2) passed

yxlm-benchmark eval-v0: 50 tasks, 119 scored items. A pass means the workflow ran on the real vendor engine and matched the expected answer on the visible data and on a hidden second dataset.

L2 pass rate by model Claude Opus 5.5 with the workflow skill 100 percent. yxlm-0.3 9B 22 percent. yxlm-0.2 9B 20 percent. yxlm-0.1 9B 16 percent. Claude Haiku 4.5 with the skill 14 percent. yxlm-0.1 4B 12 percent.
Opus plus skill is the frontier bar. yxlm-0.3 is at 22% on L2 (both runs) and 12 to 14% on terse L3 requests, where Opus scores 84%.

Read this before the chart

  • It compares systems, not models. Each Claude call carried the 72 KB workflow skill and retrieved tool cards, about 36k tokens. yxlm got the request and input schemas, about 0.5k.
  • 0.3 vs 0.2 is a tie. 18 vs 18 passes out of 100 (p = 1.0). With 50 items per level, a few points is noise.
  • What is real: 0.3 beats 0.1-4B (p = 0.008), and it learned to ask. On deliberately vague requests it asks a clarifying question 11 times in 19, up from 1 in 19 for 0.1-9B (p = 0.002). Opus: 12 in 19.
  • The numbers hold up. An independent review recomputed every leaderboard figure from the per-item results. All of them reproduce, apart from one rounding nit.

42

My AI operator.

42 is an AI agent that works for me on a Mac mini, under its own non-admin account. It reads the repos, does the work and opens pull requests. It does not merge them.

It built this page, and traced every claim on it back to a commit, a file or a URL.

  • securityPR #164 Found and fixed a critical security issue in Overlane's server. Fix in review.
  • ci greenPR #165 Cleared newly published high-severity dependency advisories that were failing CI, including a zod 3 to 4 migration kept safe under a strict CSP. 4,068 of 4,068 tests pass.
  • audityxlm-benchmark A sceptical review: re-verified 348 file hashes, recomputed every number across 9 runs, added confidence intervals and paired tests. The caveats in section 02 come from it.
  • this pagebrand-site Static, no third-party requests, every claim sourced.

AI at work

At a Jersey data and AI consultancy serving 90+ financial services clients.

  1. April 2026

    Joined as a paid intern. The firm had no AI subscription.

  2. Three months in

    Promoted at the three-month review.

  3. By October 2026

    Took a firm with no frontier AI subscription onto Claude: three people now Claude certified, and on track to become a full Claude partner.

AI workflow builder skill

30+ releases, used firm-wide every day, and deals closed on the back of it. Audited with about 85 probe workflows.

A live capital-call pipeline

One MCP call posts transactions into a fund ledger, verified against a Snowflake mirror.

An MCP gateway on Azure

Several MCP servers behind one endpoint on Azure Container Apps, with Entra ID OAuth, per-user audit logs, and a human approval gate before anything publishes to production.

Claude Certified Developer Foundation July 2026, just under 90% against a 72% pass mark.

Advanced certified on the market-leading enterprise data-prep platform.

No degree

A different education.

Costa Rica

Day Leader, Raleigh International

Three months, in 2025.

Plum Village

A month in a monastery

Silence, manual work, no phone.

Thailand

Muay Thai

Training on my gap year.