Hansel
← The village

No. 03 · The observatory

bloop

Cautionary case. COBOL-to-Java migration service that outlived its runway.

Status
Open
Case
One of eight recorded cases
The observatory in the gingerbread village
Snapshot
Scored 2026-10-05, reported 2026-10-09
Milestones
5 recorded, 2 with an exact day
Limits
Partial and unavailable dates stay as recorded text. The timeline plots exact days only.

bloop

bloop, a London YC S21 startup, converted COBOL into readable Java. A static transpiler, LLM refactoring and behavioral tests ran as fixed-price pilots. It earned at least one six-figure payment and named no customer publicly. It pivoted to Vibe Kanban in June 2025 and shut down on April 10, 2026. The founder says the technology worked; the sales clock outran the cash.

Sources:

The market it attacks

IBM says 45 of the top 50 banks value its mainframe for critical workloads. An Anthropic post cites an estimate of COBOL behind about 95% of US ATM transactions. One prospect told the founders it ran a billion lines of COBOL. TSB’s failed 2018 platform move drew a £48.65M fine plus £32.7M redress. GAO counts critical federal systems at 23 to 60 years old, costing $754M a year. bloop had yet to see a customer with tests; COBOL skills were retiring out. On bloop’s own benchmark GPT-4 scored 8.9%, against 67% on Python.

Sources:

Origin and the first wedge

Knight-Webb, from Who Targets Me, and Gordon-Hall, an ML engineer, founded it in 2021. Copilot overtook its code-search product about a month into YC. The enterprise code-search product predated GPT-4; a $12-seat version launched March 2023. Sales calls surfaced the arbitrage: LLMs write good Java and poor COBOL. Transpile mechanically, refactor the Java, use original behavior as ground truth. IBM made its own COBOL-to-Java assistant available in October 2023. The COBOL product launched March 20, 2024; it was called bloop modernise.

Sources:

The first customers

The launch carried three products; by January 2025 only Modernise remained. Its 34B model scored 10.27% on COBOLEval, ahead of GPT-4’s 8.9%. The paying job: convert a cluster touching database, file and network to Java, fixed-price. Delivery ran survey, proof-of-concept sprint, then full modernisation under an agreed matrix. bloop absorbed overruns; its first commercial codebase held 5 million lines. Buyers had already decided to migrate, so bloop sold proof, not the decision. No public source names a customer; a handful of engagements is the supportable count. It reported paying customers by June 2024 and one six-figure Canadian payer. Reach came through developer events, the GSE community, Zorse and an Accenture lab. Tests came from historic batch data, recorded green screens and LLM-written checks. A 4,096-token output cap forced method-by-method refactoring; some buyers self-hosted.

Sources:

How big it is now

It lived on 2021 funding: $125K YC money plus a $7.31M December 2021 round. Total recorded capital is $7.43M; no later round is listed. UK accounts show net liabilities climbing to £3.73M by April 2025. Average UK staff held at three to six against a ten-person YC page. Vibe Kanban reached 30,000 monthly active users and 25,000 GitHub stars. In November 2025 it was hiring two senior engineers at £100,000 each. The UK company remains legally active after the April 2026 wind-down.

Sources:

What made it hard

Expected 18 to 24 month enterprise cycles collided with a 2021 runway. Four to six staff built a transpiler for undocumented COBOL dialects. Mechanical Orchard, the rival, had $50M fresh and 80-plus staff. The Canadian deal died when its buyer-side contact was fired, during tariff turmoil. The founders had no mainframe and admitted no taste for selling to banks. Bank buyers respond to risk, and to redundant developers after migration. Some customers barred sending code to outside model providers. Batch tests needed each customer’s historic inputs and outputs. Nothing published shows VSAM or Db2 data migration. AWS made its mainframe agent generally available in May 2025, priced at zero. The founder infers AWS recoups it through decades of cloud bills, Accenture via transformations. IBM’s converter is pitched to keep workloads on IBM Z. By February 2026 the founder said off-the-shelf tools could convert COBOL. Anthropic’s February 23, 2026 guidance claimed quarters, not years; IBM fell 13%.

Sources:

The strategy

It was migration-led, delivered as an AI-run service the founder called services. The overlay products, assistant and code understanding, dropped within ten months. The plan was translation first, then maintenance and new features on the Java. Pilots and one six-figure engagement happened; no production cutover is recorded. Gartner forecasts over 70% of 2026-starting mainframe exits missing intended benefits. It also forecasts 75% of mainframe-exit vendors pivoting or ceasing by 2030. COBOLEval survives, hosted by the Zorse project; the transpiler is not public.

Sources:

Milestones

  • March 2023: GPT-4 code search launches on Hacker News.
  • March 20, 2024: COBOL modernisation products announced.
  • January 2025: Founder claims no pilot had failed to continue.
  • June 2025: Pivot to Vibe Kanban.
  • April 10, 2026: Company shuts down.

Sources:

What others can copy

Size runway to the buyer’s clock: cash should exceed 1.5 times the pilot-to-contract time. Contract phase two inside phase one: pre-agreed pricing and three named sponsors. Hire the enterprise seller early: test for a closed seven-figure regulated deal. Earn after cutover or partner with who does: over half the value lands post go-live. Do not build the moat on a model weakness: rerun your benchmark each quarter. Sell the equivalence evidence: a per-module pack the risk team signs alone. These are founder tests to run, not results already achieved.

Sources:

Open questions

Whether any conversion reached production, and the 5M-line project finished, is unpublic. The no-failed-pilots claim predates the Canadian cancellation. A systems-integrator or hyperscaler partner was never tested. Standalone migration margins against a free AWS agent are unproven. Whether a deterministic transpiler still matters is unresolved. The converter itself is absent from public repositories.

Sources:

Limits

The 18-to-24-month cycle is the founder’s expectation, not completed deals. COBOLEval measures COBOL generation, not translation quality. UK liability trends infer operating loss, never audited revenue. Enterprise-seller advice came from a whole-company retrospective.

Sources:

Exhibit 8Explore · open a case

Milestones in the bloop case

Exact-day events on a UTC timeline, one lane per case.

  • milestone

How to readEach dot is a dated milestone. Only exact dates draw dots.

NoteMonth- and year-only milestones stay in the tables, never plotted.

SourceLegacy Software Atlas. Frozen scoring snapshot of 5 October 2026. Report version 4 dated 9 October 2026.

Recorded milestones

Every recorded event for this case, in frozen order. Statuses stay as recorded.

MilestoneDateStatusSources
GPT-4 code search launches on Hacker NewsMarch 2023reportedHacker News, Launch HN: Bloop (YC S21), Code Search with GPT-4, 20 Mar 2023.
COBOL modernisation products announcedMarch 20, 2024reportedbloop blog (Louis Knight-Webb), Modernising Legacy Code with AI, 20 Mar 2024 (Wayback capture 31 Mar 2024).
Founder claims no pilot had failed to continueJanuary 2025reportedScaling DevTools (Jack Bridger), episode 117, Louis Knight-Webb from Bloop.ai, the YC startup turning COBOL into Java, page with full transcript, 2 Jan 2025.
Pivot to Vibe KanbanJune 2025reportedGitHub REST API, BloopAI/vibe-kanban repository metadata, accessed 5 Oct 2026.
Company shuts downApril 10, 2026failedVibe Kanban blog (Louis Knight-Webb), Goodbye bloop, 10 Apr 2026.

Back to the cases chapter and the full milestone table.