Miju Labs

← All buyers · Cohere

Saurabh Dash

Researcher, Cohere Labs (multimodal) · Cohere

medium confidenceTier A · Talk to firstSecond-wave labsimage · multimodal

First author of Aya Vision, Cohere's multilingual multimodal model, and the person who built its evaluation apparatus — AyaVisionBench and m-WildVision, both of which score model outputs on open-ended visual prompts by comparison. He is Cohere's most identifiable technical requester for image-and-text human judgement, though the demand is multilingual/document understanding rather than design taste.

Background

Undergraduate at IIT Kharagpur, where he led the RoboSoccer group that became the first Indian team to qualify for RoboCup SSL, then a PhD at Georgia Tech with Saibal Mukhopadhyay on efficient machine learning — quantisation, spatiotemporal learning, computationally efficient intelligent systems. He interned on Apple's MIND team in 2021 before joining Cohere.

At Cohere Labs his work spans multimodality, reinforcement learning and LLM efficiency, with publications at NeurIPS, ICML, ICLR, CVPR and ICCV. Aya Vision (May 2025) is his highest-profile output: a multilingual vision-language model released with its own benchmark suite and a synthetic-recaptioning data pipeline.

What they run now

Multimodal model building at Cohere: vision-language training data pipelines (including synthetic recaptioning), and evaluation. Aya Vision shipped with AyaVisionBench and m-WildVision, human-preference-style benchmarks for open-ended multilingual visual prompts. Around him, Cohere runs an unusually large in-house annotation operation — an Ashby department called Data Quality (Contract) with STEM, code, generalist, safety and Mandarin annotation teams — and the Command A report describes preference data built from majority vote of at least three annotators on a customised in-house platform, human pairwise evaluations using roughly 65 annotators per run, and safety annotation drawing on internal annotators plus specialist external vendors.

What they buy

Evidenced at team level: Cohere buys and runs pairwise human preference annotation (≥3 annotators, majority vote) and human pairwise evaluation runs (~65 annotators), mostly through its own contract annotation workforce, with specialist external vendors used for safety. Dash's own programmes would need multilingual raters who can judge image-grounded responses — captions, document understanding, visual reasoning — rather than aesthetic quality. Inference: he specifies rubric and rater qualifications for multimodal evaluation; he does not procure.

Career

  1. presentResearcher (multimodality, RL, efficiency), Cohere / Cohere Labs — First author, Aya Vision (2025).
  2. 2021 – 2021ML research intern, MIND team, Apple — Supervised by Hema Koppula.
  3. —Georgia Institute of Technology, Ph.D., ECE (advisor Saibal Mukhopadhyay)
  4. —IIT Kharagpur, B.Tech

On the record

Multilingual multimodality · 2025-05

Argues multimodal ability should not degrade outside English; Aya Vision targets the multilingual visual gap. source

Interviews & talks

Why it matters here

Middling for a taste-data vendor. Cohere is enterprise-first and its visual demand is comprehension, not design; it also insources annotation aggressively, which is the hardest competitive position to sell into. The one genuine hook is Europe: Cohere acquired Aleph Alpha in April 2026 and runs offices in Berlin, Paris and London, so a German vendor supplying EU-resident multilingual raters has a sovereignty argument that Cohere's enterprise customers actually care about.

How to reach

Speculation: lead with multilingual rater supply, not taste. A credible opener is European-language visual evaluation — native raters judging image-grounded answers in German, French, Italian — benchmarked against the agreement levels Cohere already publishes for its in-house pool. He is approachable as a researcher (personal site, X, GitHub) and responds to technical framing; the procurement conversation belongs with whoever runs Data Quality (Contract), not with him.

What we could not establish
  • His exact title and seniority at Cohere; his own site says only 'Researcher at Cohere Labs'.
  • Whether Cohere uses any named external vendor for non-safety annotation.
  • Whether the Aleph Alpha acquisition moved any annotation work into Germany.

Why they are on the map

First author of Aya Vision (May 2025): recaptioning data pipeline, AyaVisionBench and m-WildVision human-preference-style evals source — evidence dated 2025-05

Needs the data for a model; writes the spec.

How Cohere buys data

Runs a large in-house annotation workforce: Ashby department 'Data Quality (Contract)' with teams such as 'STEM Annotation - Engineering' and 'Code Annotation - Data Science' (Apr–Jul 2026), plus Generalist, Safety and Mandarin annotation specialists. Command A tech report (Apr 2025) says safety annotation uses 'internal annotators and specialist external vendors', multilingual annotation uses internal and external annotators, preference data uses majority vote of ≥3 annotators on a customised in-house platform, and human pairwise evals use ~65 annotators per run. Hired a 'Product Manager, Human Data' (removed Oct 2025) for 'model in the loop' annotation at scale; 'MTS, Data Analysis and Evaluation' (London, Dec 2025) designs data-collection tasks.

Modeling org (post-training, multimodal, synthetic data, evaluation, multilingual) under Chief AI Officer Joelle Pineau (since Aug 2025) and Chief Scientist Phil Blunsom; a separate Data Quality (Contract) org houses in-house annotators; Cohere Labs (research, now led by Marzieh Fadaee) publishes on multilingual/multimodal data.

Sources

  1. Personal site — Saurabh Dash (role, education, links)
  2. Aya Vision (arXiv 2505.08751)
  3. Command A technical report (annotation practice, ~65 annotators per human eval run)
  4. Cohere Ashby job board — Data Quality (Contract) department