Our work

The SAFL Timeline

. It begins in 2022 with the first scoping of India's AI safety talent and runs to the research, fellowships, and convening we do today.

Track record
2022
80+campuses mapped across India's top CS universities
Strategy

We scoped where India's AI safety talent was, and how to reach it.

We began full-time AI safety field-building in India in September 2022, founding the India AI Safety Initiative.

  • Funded by a $65K EA Infrastructure Fund pilot grant, with a further grant offer from Coefficient Giving.
  • Built on an eight-month survey across India's top computer-science universities, covering 3,000+ miles, 12 cities, and 80+ campuses.
  • Produced a four-pathway field-building proposal whose in-house research-lab pathway later became SAFL.
  • Two findings shaped what followed: career legibility beats philosophical framing, and short paid technical assignments surface stronger candidates than reading groups.
2023
600+applicants across 40 STEM universities
Fellowships

We founded India's first Alignment Research Fellowship.

It drew more than 600 applicants across 40 STEM universities, and we selected 24 fellows, the top 4% of the pool.

  • We ran a ten-person campus recruiting pod across interpretability, robustness, scalable oversight, and model evaluations.
  • We evaluated a 1,000-member pool of top-decile computer-science, physics, and mathematics students.
  • Ten papers from the fellowship are now published or under review.
  • Alumni have gone on to independent research or roles at labs including the Mila-Québec AI Institute.
The AI Safety Careers Fellowship cohort at IIT Delhi
Fellowships
20 of 22completers committed to an alignment project within five months

We ran India's first in-person AI safety careers fellowship, at IIT Delhi.

Held from August to November 2023, it adapted BlueDot Impact's AGISF curriculum to the Indian context and added capstone projects, placement support, and mentor matching, advised by FAR.AI and SERI MATS.

  • 900+ expressions of interest, 175+ full applications, and 22 completers.
  • Within five months, 20 of the 22 had committed to a significant alignment research project.
  • Six of the fourteen who applied were accepted to GCP Oxford, and one facilitator was selected for SERI MATS in Berkeley.
  • Multiple fellows turned down private-sector offers, including a confirmed ~$400K quant-firm offer, for alignment research.
2024
Research & evaluations
3NeurIPS workshop acceptances

We did technical research on evaluating AI agents, as SteadRise Alignment Research Fellows.

We evaluated how capable and reliable language-model agents are, and contributed dangerous-capability evaluations to the UK AI Security Institute.

  • Two projects accepted across three NeurIPS workshops.
  • Contributed an implementation of Humanity's Last Exam to the open Inspect Evals library.
3× NeurIPS workshops

Auto-Enhance: evaluating AI agents that improve other agents

A meta-benchmark that measures how well language-model agents can improve other AI agents. Built as SteadRise Alignment Research Fellows, it was accepted across three NeurIPS workshops.

EvaluationsLLM agentsBenchmark
Read the paper →
Summer Pilot Fellowship

LLM agent evaluations research

The research program behind the benchmark — a cohort of fellows running technical evaluations of how capable and reliable language-model agents are.

Research fellowshipEvaluations
See the projects →
UK AI Security Institute

Humanity's Last Exam → Inspect Evals

We contributed dangerous-capability evaluations to the UK AI Security Institute, including an implementation of Humanity's Last Exam added to its open Inspect Evals repository.

Dangerous capabilitiesOpen source
View the eval →
The weighted talent index that scores candidates
Talent infrastructure

ML Research Talent Map

A 50,000-profile map built from five years of ICLR, ICML, and CVPR proceedings.

  • Two-stage candidate matching using vector embeddings and a weighted talent index.
  • Lets us read the field by research area and surface strong candidates before they apply.
The multi-source recruiting data pipeline
Talent infrastructure

Recruiting Data Infrastructure

The base layer the rest of the talent systems build on.

  • Python scrapers ingest 100+ data sources, including conference proceedings, Olympiads, GitHub, Codeforces, and the 3Blue1Brown virtual career fair.
  • Produces an 80,000-record dataset at over 90% field-level accuracy.
  • Runs on Cloud SQL and Cloud Run, with Ashby ATS and LinkedIn Recruiter integrations.
A slide from the EAGxIndia field-building talk
Writing & strategy
2024field talk, EAGxIndia

We codified two years of field-building lessons in a public talk.

At EAGxIndia 2024 we set out what we had learned running programs in India, where the standard Western playbook under-delivers.

  • Career legibility beats philosophical framing.
  • Short paid technical assignments outperform reading groups for surfacing talent.
  • High-fidelity small cohorts beat large, low-touch ones.
2025
The Global AI Safety Fellowship
Fellowships
10placements at alignment and governance research labs

We built and ran the Global AI Safety Fellowship at SteadRise.

A targeted-assignment, mentorship-led model that surfaces technical and governance talent from regions Western-style fellowships under-reach.

  • We give strong candidates a meaningful technical problem upfront, rather than an eight-week curriculum, and see who can already do the work.
  • The funnel reached more than 5,200 candidates across global hubs.
  • It produced 20+ offers and 10 accepted placements at partner labs.
Researchers contributed to
Anthropic FellowshipUK AI Security InstituteFAR.AI
The AI Security IIT Delhi initiative
Fellowships
6-weekfellowship with funding, mentorship, and compute

We co-built a student-and-researcher AI safety initiative at IIT Delhi.

Started in June 2025 and modelled on HAIST and the Berkeley AI Safety Initiative, but built for the Indian subcontinent talent pool.

  • A six-week fellowship with funding, mentorship, and curated network access.
  • Mentorship and compute for advanced researchers across a technical and a policy track.
  • Active research on reasoning in language models, fairness benchmarks, corrigibility, and AI regulation.
  • Run by a roughly twelve-person core team, with IIT Delhi and the Kairos Project.
The talent graph engine connecting researcher profiles
Talent infrastructure
68,200+ML researchers and engineers mapped

We built talent infrastructure that finds safety talent before it is looking.

It is recruiting on the surface and a continuous, instrumented read on the global AI talent market underneath.

  • Maps more than 68,200 ML researchers and engineers from public signals.
  • Tells us which sub-fields are growing and which organisations are losing people.
  • A graph-based platform connects 100K+ profiles across 500+ public sources.
Placed candidates into
UK AI Security InstituteFAR.AIApollo ResearchGovAIAnthropic Fellowship
Skill-dimension fingerprints for different roles
Talent infrastructure

Talent Index & JD Fingerprinting

A data science project that standardises what strong looks like for each role.

  • Analyses 106 job descriptions across ten skill dimensions.
  • Gives a single, comparable view of role requirements and candidate-to-role fit.

Nexus

Talent infrastructure

A signal-monitoring system that surfaces experienced engineers during workforce transitions.

  • Watches public channels including Reddit, GitHub, HackerNews, and Google Custom Search.
  • Flags strong candidates entering the market quickly, so outreach happens while they are still deciding.
The two-stage candidate screening pipeline
Talent infrastructure

AI Candidate Screening Pipeline

A two-stage screening pipeline built on the Claude API and Airtable.

  • Scores enriched profiles across 20+ signals with tri-verdict logic and tier-based shortlisting.
  • Ingests CVs, LinkedIn, and GitHub, with citations back to the source.
  • Runs across more than 15 role-specific configurations.
The India AI Tracker dashboard
Governance

India AI Tracker

A dedicated tracker of the Indian AI ecosystem.

  • Covers governance discussions, stakeholder decisions, investment trends, and technological developments.
  • Follows 50+ stakeholders across government, academia, and industry.
  • Used as a reference by international organisations evaluating India as a talent source and research partner.
3 tiersleverage framework for India AI safety capital
Writing & strategy

We published a framework for India-specific AI safety grantmaking.

It orders deployment by marginal leverage rather than gross spend.

  • Tier 1 compounds: evaluations capacity, alignment research at top Indian institutions, and a regranting function.
  • Tier 3 mostly does not: general convenings, regardless of dollar size.
  • It sits inside a wider approach: judge programs by what would have happened anyway, and be honest about which models transfer to India.
September 2025Secure AI Futures Lab is founded.

Work before this point was led by the same founders under Axiom Futures and Impact Academy, before SAFL was formed.

2026
SAFL's workshop at the India AI Impact Summit 2026
Convening
50+government, academia, and industry stakeholders convened

We ran SAFL's Hardware-Rooted Sovereignty Workshop at the India AI Impact Summit 2026.

Held at Bharat Mandapam in February 2026.

  • Featured Prof. Stuart Russell and Eileen Donahoe, with a live demo of Lucid Computing's hardware-rooted sovereignty stack.
  • We appeared on stage with India's IT Secretary, S. Krishnan.
  • The agenda covered hardware-layer verification, India's DPDP Act, and compute supply chains as the anchor for AI trust-fabric standards.
Pre-summit Trustworthy AI panel in Singapore
Convening

Trustworthy AI convenings

We built the rooms where technical depth and government reach meet, across Singapore and New Delhi.

  • Convened the official pre-summit panel for the India AI Impact Summit at SUTD in Singapore on 25 January, during AAAI-26 week.
  • Speakers spanned India, Singapore, and Australia, including MDDI, the NUS AI Institute, IIT Madras (CeRAI), UNSW, AIDXTech, and FAR.AI.
  • Hosted a baithak-style academic reception near Bharat Mandapam with academics from IITD, IITM, IIITD, IIITH, UNSW, MBZUAI, and ASU.
Where this leads

We unlock research expertise and enable institutional capacity to steer the most powerful technology of our time.

SAFL pursues parallel interventions in AI governance and impacts research, expert talent advancement, government engagement, and ecosystem support infrastructure. Our foundational work received support from Schmidt Sciences and the Future of Life Institute, among others. We work across tech hubs in India and South Asia, including New Delhi, Bengaluru, Chennai, Hyderabad, Mumbai, and Singapore.

Research

Technical research to help India and other low- and middle-income countries navigate advanced AI.

Careers

Support for early- and mid-career talent moving into AI safety, evaluations, and governance work.

Events

Convening government, academia, and industry across India and South Asia to grow the field.

SAFL Hubs

Physical and hybrid hubs across Delhi, Bengaluru, Chennai, and Hyderabad.

See what we are working on now →