Miju Labs

← All buyers · Mistral AI

Khyathi Raghavi Chandu

AI Research Scientist, Mistral AI (Palo Alto) · Mistral AI

medium confidenceTier A · Talk to firstSecond-wave labsmultimodal · image · audio

A multimodal research scientist at Mistral and a core contributor on Magistral (reasoning/post-training) and Voxtral (audio), with a research history in grounded multimodal generation and in how models are evaluated. She is the kind of person who writes the spec for what human annotators are asked to judge, rather than the person who signs the contract.

Background

PhD at Carnegie Mellon's Language Technologies Institute, advised by Alan W. Black, Eric Nyberg and Yonatan Bisk, on grounded multimodal long-form generation; named a UC Berkeley Rising Star in 2020. She was a research scientist at Meta AI and then on the Allen Institute for AI's Mosaic team under Yejin Choi, where she shared best-paper awards at ACL 2024 for OLMo and Dolma — two projects defined by radical openness about training data.

She joined Mistral as an AI research scientist working on multimodal LLMs, based in Palo Alto rather than Paris. Her community record is heavy on evaluation and inclusion work: D&I chair at ACL and NAACL, Widening NLP organiser, area chair at COLM and AACL.

What they run now

Model development from 3B to 123B parameters at Mistral, with named contributions to Voxtral (open audio-language models) and Magistral (reasoning). Mistral's Ministral 3 report describes curated multimodal SFT data and a pairwise reward model; researchers in her position are the internal customers of the Human Data Annotation team — they define the capability a campaign should improve and the criteria annotators apply. Her AI2 background means she is unusually fluent in dataset documentation and provenance, which cuts both ways for a vendor: she will ask harder questions about how data was collected and who collected it.

What they buy

Inference. As a multimodal core contributor she plausibly specifies evaluation sets and preference data for vision and audio capabilities, and reward-model training pairs. Her AI2 lineage (OLMo/Dolma) suggests a bias toward open, well-documented data and possibly toward building rather than buying. No public statement ties her to any vendor.

Career

  1. presentAI Research Scientist (multimodal LLMs), Mistral AI — Based in Palo Alto; core contributor on Magistral and Voxtral.
  2. —Research Scientist, Mosaic team, Allen Institute for AI — OLMo and Dolma; best-paper awards at ACL 2024.
  3. —Research Scientist, Meta AI
  4. —Carnegie Mellon University (LTI), Ph.D., grounded multimodal long-form generation
Why it matters here

Medium. She is a credible technical champion inside Mistral for multimodal human data, and she is reachable in the normal academic way (papers, conferences, X). But she sits in Palo Alto, works on audio and reasoning more than visual design, and is not a budget holder — so she is a route to internal advocacy, not a sale.

How to reach

Approach as a researcher, not a vendor. Speculation: the opening that would work is an evaluation question rather than a capability pitch — e.g. how Mistral handles rater disagreement on subjective multimodal quality, and an offer to share a small annotated set with agreement statistics attached. Meeting her at a conference (ACL/COLM/NeurIPS) is more natural than a cold email, and anything sent should be documentation-first given her OLMo/Dolma background.

What we could not establish
  • Start date at Mistral and whether her remit includes image/visual quality at all.
  • Whether she has any say over external annotation vendors.
  • Her exact internal title beyond 'AI Research Scientist'.

Why they are on the map

Core contributor on Magistral (RL/post-training) and Voxtral (2025); background in multimodal human-preference evaluation — likely specifier of eval/preference data source — evidence dated 2025-07

Shapes what 'good data' means.

How Mistral AI buys data

Runs an in-house 'Human Data Annotation team within the Science organisation' (Data Annotation Quality Specialist, Paris, permanent, posted ~mid-2026). Repeated TPM hires for 'Human Data Annotation' (incl. a Code variant for Paris/London/Palo Alto) whose remit is 'contractor teams and vendor relationships for data collection campaigns with subject matter experts' plus throughput/budget/quality dashboards. A 'Data Annotation Technical Program Manager' role was already being advertised in May 2024. Ministral 3 report (Jan 2026) describes curated multimodal SFT data and a pairwise reward model but names no vendors.

Science org under co-founder/Chief Scientist Guillaume Lample contains the Human Data Annotation team (in-house STEM-PhD annotators + TPMs who manage external vendors/contractors). Research papers use large alphabetical author lists with a small 'core contributors' block; multimodal group formerly led by Devendra Chaplot (left; at xAI since Mar 2026).

Sources

  1. Personal site — Khyathi Raghavi Chandu (Mistral, career, links)
  2. LinkedIn — 'Ph.D, Multimodal LLMs @ Mistral AI'
  3. Voxtral technical report (Mistral, 2025-07)
  4. Magistral technical report (Mistral, 2025-06)