First-author builder spanning endoscopy AI and ML systems: ~100 fps real-time inference, a 5-level-verified Zig stack
2
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Evolve holds no verified publication count for this name — only two 2026 arXiv preprints — so output volume and venue mix cannot be scored. This axis would measure total papers, first-author share, and yearly production trajectory.
Impact—
No citation counts or venue-level influence data are available in this snapshot, so impact cannot be scored. This axis would measure citations, citing works, and real-world uptake of the woma and Numbat artifacts.
Focus0/100
The focus dimension scores 0 because the raw field distribution is perfectly split: medical image analysis (endoscopy computer vision and ML systems engineering) at 0.5 and software engineering (ML systems) at 0.5 across the two cataloged preprints. woma drives the medical side; Numbat drives the systems side. Rather than unfounded drift, this reads as a deliberate bridge between clinical computer vision and reproducible ML infrastructure — but two papers are a thin basis for judging true focus.
Collaboration14/100
The bibliometric record shows exactly 1 distinct co-author, and the collaboration dimension scores 14. Lan Dang is the sole listed frequent collaborator, appearing across a two-paper, two-field output — endoscopy computer vision and ML systems engineering. Both preprints are first-authored without a corresponding-author flag, suggesting a compact, engineering-driven working pattern. With so small a network, external validation, multi-site datasets, and clinical co-authorship remain untested; widening the collaborator base would materially change this axis.
Activity—
Only a single publication year (2026) is present, with no multi-year record, so recent activity cannot be scored. This axis would measure output recency and sustained publishing cadence over time.
Reproducibility55/100
The reproducibility dimension scores 55, the mean over 2 papers both rated medium readiness. The underlying practice is unusually explicit: woma is pre-registered and reports honest misses, while Numbat claims five-level-verified training parity with YOLOv8m (0.4956 vs 0.502 mAP). Both carry the same caveat — woma's shipped weights were chosen on the test record, and Numbat is single-seed with a closed-source framework — which keeps the score mid-range rather than high.
Research summary
Thang Tran is an early-stage researcher whose visible record — two 2026 arXiv preprints, both first-authored — sits at the intersection of medical image analysis and ML systems engineering. The first, woma, introduces a real-time foundation model for gastrointestinal endoscopy built with self-supervised learning, plus fine-tuned variants for colonoscopy and gastroscopy tasks such as polyp detection; Evolve's review notes pre-registration, honest reporting of missed results, and inference faster than a TensorRT baseline at roughly 100 fps, while flagging that shipped weights were selected on the test record. The second, Numbat, is a self-contained machine-learning stack written in Zig, reporting five-level-verified training parity with YOLOv8m (0.4956 vs 0.502 mAP), though evidence rests on a single seed and a closed-source framework. Bibliometrically the profile is narrow: exactly 1 distinct co-author (Lan Dang), a collaboration score of 14, a focus score of 0 reflecting an even 0.5/0.5 split across the two fields, and a reproducibility score of 55 averaged over 2 papers, both rated medium readiness. Output, impact, and activity axes lack sufficient data to score. The name is resolved from public bibliographic data; no institution, country, or ORCID is on file.
Collaboration pattern
The recorded network is minimal: 1 distinct co-author, giving a collaboration score of 14, with Lan Dang as the only frequent collaborator. Both 2026 preprints are first-authored, non-corresponding, and span two fields — endoscopy computer vision and ML systems engineering — suggesting a tight two-person, engineering-led collaboration. No clinical, multi-site, or industry co-authors appear yet; co-authorship with domain clinicians and independent replication groups would be the natural next step.
Research keywords
Highlights
Real-time endoscopy foundation model, pre-registered with honest misses
A self-contained Zig ML stack verified to training parity with YOLOv8m
A reproducibility-first working pattern across two fields
Both 2026 preprints show the same engineering discipline: explicit verification (pre-registration for woma; five-level parity checks for Numbat), transparent reporting of limits, and medium reproducibility readiness with a mean score of 55 over 2 papers. This thread is the connective tissue between the endoscopy and ML-systems lines of work.
Timeline
- 2026First-authored the Numbat preprint on arXiv: a self-contained machine-learning stack built in Zig, reporting five-level-verified training parity with YOLOv8m (0.4956 vs 0.502 mAP).
- 2026First-authored the woma preprint on arXiv: a real-time foundation model and fine-tuned models for gastrointestinal endoscopy, with self-supervised pretraining covering colonoscopy, gastroscopy and polyp detection.
- 2026Closed a two-paper, two-field first-author year with a single recorded co-author (Lan Dang), a mean reproducibility score of 55 across 2 papers, and an even 0.5/0.5 field split between endoscopy computer vision and ML systems engineering.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Bibliometric scores are derived from public bibliometric aggregation (OpenAlex-type sources); this snapshot contains no citation counts, venue records, or h-index.
- Profile is name-resolved from public bibliographic data: no institution, country, or ORCID is attached, so identity confidence is limited.
- The Evolve catalog holds only two 2026 arXiv preprints for this researcher; output, impact, and activity dimensions are unscored for insufficient data.
- Performance claims (~100 fps; 0.4956 vs 0.502 mAP) are quoted from Evolve's own paper reviews and are not independently verified.