Hi, I’m Anuradha.
I’m curious about
how machines think
& how people live.

I’m a grad student in Business Analytics & AI at Stevens. I spend most of my time building LLM tools and figuring out how to test them properly. The rest of the time I’m usually somewhere new, eating something I can’t name.

Currently wondering about golden-set evaluation

AIRS Research Fellow, Stevens Institute for Artificial Intelligence · Provost Master’s Scholar · M.S. Business Analytics & AI

See what I’ve built → Start with a question
Anuradha Patil

Questions I keep chasing

These are questions I keep running into. Pick what you think, then see what I think. If we disagree, that’s probably a good conversation.

01 · AI

Can an LLM read a million papers without pretending to understand them?

It can read them, but it can’t tell you which ones matter. What worked for us was attaching a confidence score to every claim and making sure a person signed off before anything went into the knowledge base.

← next question
02 · DATA

Which usually breaks an ML project first?

The pipeline, in my experience. Duplicate records, missing IDs, dates in three formats. I’ve spent far more hours cleaning data than choosing models.

← next question
03 · ANALYTICS

A dashboard says sales rose 12%. What's my first question?

Compared to what? Last month, last year, the same week before a promotion? Most of analytics is not accepting the first number you’re shown.

← next question
04 · AGENTS

Should an AI agent act before it checks its own work?

Validate first. An agent that acts without checking itself is just fast, and fast in the wrong direction gets expensive.

← next question
05 · BUSINESS

What did nine years in manufacturing teach me about AI?

The process decides. Every tool I’ve worked on either got used or didn’t depending on how it fit into someone’s actual day, not on how good the model was.

← next question
06 · GENAI

RAG or fine-tuning — which first?

RAG, almost always. If the model needs new information, give it the information. Fine-tuning is for changing how it behaves, and that comes up less often than people expect.

← next question
07 · EVALUATION

How do I know an AI product actually works?

You run it against examples that people have already labeled by hand. I’ve been surprised by the results more than once, usually in the humbling direction.

← next question
08 · OPEN

What would I build if the model were free and instant?

I don’t know yet. If you have an idea, I’d like to hear it.

← next question

Looking back at these, most of them end up being about whether people will trust the output. That seems to be the real problem.

How I think about AI

viewBox="0 0 200 44">✓The model is one step.Retrieval, the output schema, validation, logging, and the person who reviews the result all matter just as much. Writing the prompt is usually the quick part.
eval set > vibesTest before you believe.I keep a set of hand-labeled examples and run every change against it. Demos look good. The eval set tells the truth.
human hereKeep a person at the decision.Not as a backup, but as the main step. In my experience, tools that show their evidence get used, and tools that hide it quietly get ignored.
{"confidence": 0.91,"status": "verified"}Structured output over prose.Once I forced the model to return a fixed schema with a confidence field, the rest of the pipeline could actually rely on it.
Agents should check their own work.Draft, validate, retry if needed, then act. I’d rather have a slower agent that’s right.
viewBox="0 0 200 44">operationsAIMost of my AI ideas didn’t start in AI.They came from a factory floor, a procurement meeting, or a class I was grading. The technology came later.

Things I’ve built to find out

🐝

ResearchBee

A research assistant that screens papers, pulls out structured fields with confidence scores, and sends everything to a person for review before it’s saved. 2.3M+ articles indexed so far.

PythonDuckDBLLM APIsFlask
⚑

ShopPilot — a shopping agent built for agentic commerce

A retail shopping assistant that calls tools instead of guessing. It refuses off-topic requests, can’t invent products, hands off to a human when it should, and logs usage, cost, and feedback to a dashboard. Tested on a 43-case golden set. I also wrote the PRD and launch memo, because that’s the job.

OpenAI APIFastAPIDuckDBGuardrailsHuman-in-the-loopTelemetry
⚖

AgentBench-Lite — honest scores for tool-using agents

A small harness for scoring agents and models on task success, tool-call budget, latency, and cost, across providers. I built it because I kept picking models based on vibes and wanted numbers instead.

Agent evaluationMulti-providerBenchmarkingPython
▤

EventEdge — structured outputs over messy disclosures

Takes SEC filings and earnings-call transcripts and turns them into validated structured records using the Anthropic API, then shows them on a dashboard. Every field is checked before it lands.

Anthropic APIStructured outputsPydantic validationDashboard
◎

FitRec — a fashion recommender from someone who made the clothes

A two-tower recommender for fashion: FAISS for candidate retrieval, LightGBM for re-ranking, evaluated offline on recall@k and NDCG. Built for an industry I spent years in.

PyTorchFAISSLightGBMRecall@k · NDCG
▦

LoomLab — generative textile design meets pricing

Generates textile patterns from a prompt, tiles them for repeat printing, creates colorway variants, and estimates the cost to make the garment.

Generative modelsNext.jsPythonPricing logic

Code for most of these is on github.com/anuradhapatilAi ↗

Three lenses, one habit

I’ve worked in three very different settings. The question I ask is the same in all of them: how do decisions actually get made here, and where do they go wrong?

AI

What I do now.

  • RAG, agents, structured LLM pipelines
  • Golden-set evals, LLM-as-judge
  • Model selection on accuracy, latency, cost
  • Guardrails, human-in-the-loop, governance

Manufacturing

Where I learned how much a bad handoff costs.

  • Product development for 100+ global brands
  • Built forecasts and tracked KPIs for planning
  • Process redesign: dev time down 42%
  • Cross-functional product teams

Operations & e-commerce

Where I learned to check whether people actually use what you build.

  • Workflow automation across 25+ partner sites
  • SQL + Python: delays down 20%
  • Supplier and procurement AI tools: costs down 11%
  • Dashboards for time saved and adoption

The habit carried over: find where the decision breaks, fix the handoff, then go back and see if anyone is still using it.

Some things I’ve written

INSIGHTS · AI & BUSINESS ANALYTICS · 5 MIN READ

The Spreadsheet That Lied to Me (Politely)

A demand forecast, a very clean chart, and the meeting where I learned a model can be accurate and still wrong.

Years before I wrote any machine learning code, I sat in a planning meeting looking at a forecast everyone liked. The chart was smooth and the error bars were tight. It told us to cut a product line that, as it turned out, was about to have its best season in years. The forecast wasn’t wrong, exactly. It was answering a question none of us had asked.

I still think about that meeting. A forecast that fits last year well is assuming next year will look like last year. Ours had picked up a promotion schedule, a factory shutdown, and one unusually cold winter, and treated all three as the trend. The model understood the past fine. We hadn’t thought hard enough about what the past meant.

What fixed it wasn’t a better algorithm. It was asking a few plain questions before the chart went on a slide. Compared to what? Over what period? What happened in that data that won’t happen again? A lot of analytics, I think, is just not accepting the first number until you understand the story under it.

Machine learning didn’t change that for me; it just raised the stakes. A gradient-boosted model can be very confident about a pattern that stopped existing. So now anything I ship comes with a short list of what the model can’t know. It makes the demo a little less impressive. It makes the decision better, which is what we were there for.

Anuradha standing outside 40 Wall Street in New York
JOURNEYS · NEW YORK · 4 MIN READ

Standing Still on Wall Street

A street that runs on numbers, and a few minutes of standing still on it.

Everyone on Wall Street is moving. That’s the first thing you notice. The second thing you notice is that you’ve stopped. I stood outside a building with a very large name on it and watched a bike courier, a security guard, a woman on her phone, and a tourist pass through the same few square feet without looking up. I was the only one not going anywhere.

I grew up around markets that ran on different clocks. A sugar market in Sangli. A textile floor in Ahmedabad. A factory gate in Vapi where the shift change was the most reliable thing in town. Wall Street felt like the same thing in a better suit. Somewhere in those buildings, someone was still deciding what a number meant.

What stayed with me was how little of it you can see. Enormous decisions happen on that street and none of them are visible from the sidewalk. That’s true of the models I work on too. The part that matters is in the assumptions, the validation step, the one field that says how sure the model is. Nobody sees that part either.

I took the photo, put my phone away, and stayed a few minutes longer than I needed to. It seemed like a reasonable way to understand a place that never stops: stand still on it for a while.

INSIGHTS · AI PRODUCTS · 5 MIN READ

The Most Important Button in an AI Product Is “Why?”

Nobody asked for it. Everyone used it. What one small feature taught me about AI products people actually keep using.

The first AI feature I helped ship did exactly what the spec said. It looked at the data, made a recommendation, and showed a big button to apply it. In testing, people liked it. Once it was live, they mostly ignored it. The recommendations weren’t bad. Most were good. But nobody could tell the good ones from the bad ones, so they didn’t trust any of them.

We added a small link next to the button that said “Why this?” It opened a plain panel showing the three signals the model had used, the one it had ignored, and how confident it was. It wasn’t clever and it took about an afternoon. Usage went up more than it had from any model improvement that quarter.

That changed how I think about AI products. The model matters, but the product is everything a person can do around it: see the reasoning, disagree with it, undo it, and tell it something it didn’t know. If there’s no way to say no to the feature, people just stop using it.

So now I try to work in this order: what does the person need to see to trust this, what can they change, and how do they undo it. The model comes after that. It’s the easiest part to swap out and, oddly, the hardest part to get anyone to believe.

Working on next: RAG or fine-tuning, which first? · Why an agent should check its own work · A/B tests that tell the truth

Where I’ve worked

✦

AI Research Fellow — AIRS Fellowship, Stevens Institute for Artificial Intelligence

Jun – Aug 2026 · Hoboken, NJ

◆

Teaching Assistant / Student Grader — Augmented Intelligence & GenAI; Managing Data Resources, Stevens

May 2026 – present

◈

AI Product Specialist — AI startup, India

Jul 2024 – Jul 2025

⬡

Business Operations Manager — Transformation & Analytics, tech startup, India

Jul 2019 – Jun 2024

▣

Head of Product Development — manufacturing, India

Jun 2014 – Jun 2017

Internships — manufacturing & operations, India

Banswara Syntex Ltd.Jan – May 2014 · Graduation project
Asarwa MillsDec 2012 – Jan 2013
Raymond Ltd. (Silver Spark)May – Aug 2013
Alok Industries Ltd.Jun – Jul 2011
Parivartan IndustriesJun – Jul 2010

Where I learned to ask

◆

Stevens Institute of Technology · Hoboken, NJ

Master of Science, Business Analytics & AI · GPA 3.9 / 4.0 · Expected May 2027 · Provost Master’s Scholar

Business Applications in AI Systems DevelopmentIntroduction to Deep Learning and Business ApplicationsIntroduction to Python for Financial ApplicationsExperimental Design IIMarketing AnalyticsFoundation of Business AnalyticsManaging Data ResourcesManagement of AI TechnologiesData Analytics and Machine LearningMultivariate Data AnalysisOptimization and Process AnalyticsAugmented Intelligence & Generative AI
▤

National Institute of Fashion Technology (NIFT) · India

Master of Technology · First Class · 2014

◎

Shivaji University · India

Bachelor of Technology · First Class · 2012

Conferences & events

A few from this year.

Nov 4, 2026 · Hoboken, NJSIAI Fall Seminar & AIRS Fellowship Poster Session

Presented my AIRS research poster. The afternoon opened with a talk by Dr. Siddhartha Sen of Microsoft Research on using AI to model human behavior, which I’m still thinking about.

2026 · St. Pete Beach, FLSouthern Management Association (SMA) 2026 Annual Meeting

I was selected to attend through Stevens. A week of talks on evidence-based decisions and research across disciplines, and a lot of conversation about where AI fits in strategy and leadership. The thing I took home: good decisions start with the right questions.

2026 · New York, NYAWS Summit New York 2026

Keynotes and sessions on agentic AI, enterprise AI, cloud security, observability, and real-time analytics. Spent most of the expo floor at the AWS, OpenAI, Anthropic, NVIDIA, Snowflake, and Datadog booths asking too many questions.

Right now

BuildingAn evaluation harness for tool-using agents. Small tasks, honest scores, runs I can repeat.
LearningExperimental design. Mostly learning when an A/B test is telling the truth and when it’s just noise with a p-value attached.
ReadingCo-Intelligence by Ethan Mollick, AI Engineering by Chip Huyen, and The Alignment Problem by Brian Christian. Slowly, and usually two at a time.

About me

Anuradha Patil on Wall Street, New York40 Wall Street, one Tuesday

I’m a Business Analytics & AI master’s student at Stevens, an AIRS Research Fellow at the Stevens Institute for Artificial Intelligence, and a Provost Master’s Scholar. Before this I spent about ten years in product and operations roles, in manufacturing, at a tech startup, and most recently at an AI startup in India.

I’m aiming for AI product work. What I like doing is taking an idea from the first conversation with users through to a tool they actually use: writing the requirements, comparing vendors and models, setting up the evals and guardrails, rolling it out, and then checking whether anyone kept using it. I can write the code in between, which helps. Business taught me how organizations really work, analytics taught me to question numbers, and AI is where those two finally meet.

AIRS Research FellowStevens Institute for Artificial Intelligence · Summer 2026
Provost Master's ScholarM.S. Business Analytics & AI · Stevens Institute of Technology
Teaching Assistant / Student GraderMIS 630 Data Management · BIA 662 Augmented Intelligence & GenAI

Selected projects

ResearchBee (LLM research platform, 2.3M+ articles) · AgentBench-Lite (agent evaluation harness) · FitRec (two-tower fashion recommender) · StyleCast (demand forecasting) · LLMTrace (LLM observability) · github.com/anuradhapatilAi ↗

Areas of interest

Artificial Intelligence & Generative AILarge Language ModelsMachine Learning & Predictive AnalyticsProduct Analytics & Data StrategyBusiness Intelligence & Decision Science

If you’re working on something in AI or analytics and want to compare notes, write to me at [email protected]

Have a question I haven’t thought of? I’d like to hear it.

Email me → LinkedIn GitHub Google Scholar