What was built

• Mobile multimodal messaging for health education
• Monitoring dashboards for last-mile care delivery
• Gamified rewards for community health workers
• Multilingual Q&A for reproductive health (1000 Q&A dataset)

Where it went

• RCT with 100 families in Jalgaon, Maharashtra over 2 years
• Expanded to 250 families in Bangladesh for menstrual health
• Presented at Psychology of Technology Conference in DC
• Technical evaluations with Cohere (co-authoring publication)

What was built

• A system that verifies claims influencers make in their online videos across Instagram
• Focused on health and finance sectors, where misleading claims cause real harm
• Automated claim extraction from video transcripts + fact-checking pipeline

Where it went

• Presented to the head of Jagran New Media, who was also head of the International Fact-Checking Network (IFCN)
• This person worked collaboratively with the team for a year
• Presented at the India AI Summit to Ashwini Vaishnaw (Minister of MeitY)
• The Jagran head left his organization to launch a startup around this product

The problem I see

• Students publishing at low-quality journals because it seems easier
• To everyone outside your college, more papers at bad venues = less credibility, not more. You are self-selecting into a community that does not care about research but about posturing
• The bar for NeurIPS, ICML, and AAAI workshop papers is genuinely achievable in 4-6 months with a good research problem

What matters more than papers

• Write a really good technical blog post and release a well-documented library. LLMs will use your library without users even knowing, and you can cite that as impact
• This is a day and age of builders. If you are not building, ask yourself why. It is clearly not technology stopping you
• Research is about learning new things about the world, not about writing papers. The point is exploration and genuine curiosity

What the data says

• 40% of jobs globally exposed to AI, 60% in advanced economies (IMF, 2024). Exposure is not displacement, but it is task redesign
• AI increased productivity 14% in customer support (NBER) and 37% faster writing tasks (SSRN). Biggest gains for less experienced workers
• Heavy AI use makes junior developers less capable of supervising AI effectively (Anthropic)
• Value capture goes to integrators and infrastructure, not model builders. Foundation models are commoditizing

What I learned interviewing

I interviewed at OpenAI, Anthropic, and DeepMind for data science and research engineer roles. Here is what I learned:

• Debugging is the new entry-level skill. Programming is assumed, not differentiating. The bar has shifted to debugging code and planning architecture
• LeetCode-style interviews are being replaced by debugging exercises on platforms
• What differentiates you: systems thinking, the ability to plan, and building things that work in production

Project pitch (1 page, 4 questions)

1. What is the idea? (2 sentences max)
2. Why is it important? (2 sentences max)
3. What have others done and how is this different? (review 4-6 papers)
4. What experiment will highlight this difference?

Methodology from Rajesh Ranganath (NYU), Dan Fu and Jennifer Widom (Stanford)

Collaborative process

1. First meeting: write a set of questions together
2. Spend time reading and reviewing papers! Refine your research question before you start looking for answers.
3. Meet 1-3 times/week, iterate for 12-36 weeks
4. First authors handle the hardest subtasks + organizing
5. Last author does mentoring, feedback, and guidance including several 1-2 hour live editing walkthrough

What exists and why it falls short

• Keyword matching misses relevant posts using different words for the same concept
• Semantic search returns too many vaguely related results
• Social media text is short and noisy, so standard query expansion adds more noise than signal

Research questions

• How do you evaluate retrieval quality on social media data where no gold-standard relevance set exists?
• Can you build better query expansion for non-English languages where training data is sparse?
• How does retrieval precision change across platforms with different post formats?

Current state of the art

• Traditional topic models (LDA) struggle with short text
• LLMs generate labels but repeat themselves: "IndiGo Flight Disruptions" appeared in 28 out of 53 labels
• We solved this using embedding geometry to guarantee 100% unique labels

Research questions

• Can you improve clustering for multilingual data where embeddings are weaker?
• What is the right way to evaluate whether a topic label is "good"?
• Can you do theme discovery in real-time as posts arrive?

What is still hard

• Getting agents to pick the right tool for ambiguous queries
• Evaluating correctness (not just whether it ran without errors)
• Preventing hallucination when agents produce charts people trust

Research questions

• Can you build specialized agents for specific journalism tasks?
• How do you design evaluation benchmarks where policy compliance conflicts with task completion?
• What happens when you red-team an agent with prompt injection?

Real impact from student work

• 600 pages, 95M views, 500K posts analyzed on Meta
• 4,500 Telegram channels spreading pro-Russian disinformation identified
• Accepted at Stanford T&S and Underground Economy Conference

Research questions

• Can you detect coordination across platforms, not just within one?
• How do you distinguish organic consensus from manufactured coordination?
• Can temporal GNNs scale to millions of nodes in near-real-time?

Why this matters for Arbiter

• Partners in Kenya, India, Mongolia, Bangladesh need analysis in their languages
• LLMs are overconfident in languages where they perform worst (Nature Sci. Reports, 2026)
• Even small annotated datasets are publishable contributions

Research questions

• Does translate-then-retrieve or retrieve-in-native-language work better for RAG?
• Can you build a code-mixing-aware sentiment analyzer for Hindi-English?
• How does Arbiter's clustering degrade on non-English posts?

Why this is an engineering problem

• Building usable audit tools requires good engineering, not novel ML
• Exactly the kind of work undergraduate engineers are good at
• FAccT 2026 lists "audits and assurance testing" as a focus area

Research questions

• Can you build an open-source tool that automates data schema documentation for auditors?
• How would you test a recommendation algorithm for differential treatment across user profiles?
• What does an "inspectability API" for Arbiter look like?

What is broken

• Detectors trained on one generation model fail on newer models
• False positives on edited authentic content create legal problems
• Text-based AI detectors cannot reliably distinguish human from AI text (Nature Communications, 2025)

What you could build

• Benchmark AI-text detectors on code-mixed content
• Test false positive rates on edited journalistic photos
• Build a C2PA content provenance tracker

What you could build

• Measure latency gap between note creation and viral peak
• Compare note quality on polarizing vs. non-polarizing topics
• Test LLM-augmented note-writing

Key references

• Beyond Community Notes (arXiv)
• Crowdchecking Curb Misinfo (ISR, 2024)
• Trust-Consensus Paradox (ISD)

What you could build

• Simulate misinfo spread through WhatsApp groups during an election
• Validate simulation predictions against fact-checker databases
• Test proposed moderation policies in simulation before deployment

SimPPL connection

• Social network simulation at ICML '23
• Marketplace design at IC2S2 '24
• Transparency regulation at ACM DGO '24

What we look for

• Students who want to build real products that real people use
• Comfort with Python or JavaScript (we teach the rest)
• Curiosity about how information spreads online
• Willingness to read papers, run experiments, and iterate

What you get

• Co-authorship on publications at top venues
• Your code deployed and used by journalists in 8 countries
• Mentorship from researchers at MIT, NYU, Oxford, and BU
• Access to Overleaf, Zotero, compute, and real datasets