Ivy Chen
@polycount-ivyRegular
Builds evaluation suites for document AI. Messy receipts are my favourite benchmark.
- Document AI
- OCR
- Test suites
Vancouver, Canada · English, Chinese, Mandarin · America/Vancouver
Pinned
receipts-ocr-gauntletOct 19, 2026Outcome: 9 tasks across printed, handwritten and crumpled receipts. Hidden tests, rule-based grading.Test suite · v2.1Independently reproduced
- Establishes
- Another member re-ran it and matched the reported results.
- Doesn’t establish
- Does not establish results on other hardware.
- Disclosure
- Independent work
Runs
16 verified · 5 on BenchClub cloud
ModelWhereSpeed
gemma-3-4b-it Q4_K_MNov 3, 2026Apple M3 Max · 64 GB92.3 tok/s
Qwen3.5-35B-A3B Q4_K_MOct 31, 2026L40S · 48 GBCloud201.4 tok/s
GLM-4.7-Flash Q4_K_MOct 26, 2026L40S · 48 GBCloud201.2 tok/s
Mistral-7B-Instruct-v0.2 Q4_K_MOct 26, 2026RTX 4070 Ti SUPER · 16 GB109.4 tok/s
Phi-3.5-mini-instruct Q4_K_MOct 24, 2026L40S · 48 GBCloud260.3 tok/s- Qwen3-4B Q4_K_MOct 23, 2026L40S · 48 GBCloud249.7 tok/s
- Phi-3.5-mini-instruct Q4_K_MOct 19, 2026Apple M3 Max · 64 GB103.3 tok/s
- Qwen3-30B-A3B-Instruct-2507 Q4_K_MOct 10, 2026L40S · 48 GBCloud187.2 tok/s
PassedFailedNot supported
Test suites
invoice-fieldsOct 4, 2026Outcome: Suite designed for a document-AI startup's launch.Test suite · v1.0Delivery acceptedPaid collaboration
- Establishes
- The client accepted the delivery against the agreed brief.
- Doesn’t establish
- Does not establish audience reach.
- Disclosure
- Paid collaboration
Articles
Why OCR benchmarks lie about receiptsSep 27, 2026ArticleSource linked
- Establishes
- The link points to this creator's published work.
- Doesn’t establish
- Does not establish the numbers in it.
- Disclosure
- Independent work
Accounts, metrics and works carry separate evidence. Updated Nov 2, 2026. Public JSON