raghavupadhyay.org / v7Revised 2026-09Open to full-time AI / ML Engineer roles · USAuthorized to work in the U.S. on F-1 OPT — no sponsorship needed

Raghav Upadhyay

AI / ML Engineer — Research Collaborator, University of Arizona

Recent M.S. Data Science grad (May 2026) looking for AI / ML Engineer roles — hands-on with LLMs, retrieval systems (RAG), and computer vision. Based in West Lafayette, IN; open to relocating or remote.

Also searching under: ML Engineer · Applied Scientist · NLP Engineer · AI Research Engineer

“I build retrieval systems that cite their sources. A CI gate blocks the deploy when they stop.”

Abstract

I'm Raghav Upadhyay. Most people can get an LLM to answer a question. My interest is the part after that: proving the answer was grounded, and catching it automatically when it isn't.

That's what AskMyDocs is. It splits retrieval 60/40 between a vector index and BM25, reranks with a cross-encoder, and grades every answer with an LLM judge in CI: faithfulness below 0.70 or citation rate below 0.80 and the pipeline fails. It's deployed and public on Hugging Face Spaces. The same instinct produced my study of LLM-generated social networks, now under review at NeurIPS 2026: 192 generated networks across four cultures, four languages and three model tiers, measured rather than eyeballed.

That extended into vision-driven robotics: an e-puck robot solving a Webots maze from camera pixels alone, where the useful result was a negative one: the detector keyed on absolute brightness broke under dimming, and the ones keyed on relative structure didn't. That finding is now the seed for my work as a Research Collaborator at the University of Arizona, under Prof. Eungjoo Lee, asking whether a vision-only robot can tell when its own visual evidence has stopped being trustworthy.

I completed my M.S. in Data Science at the University of Arizona in May 2026 and I'm looking for AI / ML engineering and research roles in the US.

2. Method (Approach)

Three steps, in order. Skipping the second is how the first stops meaning anything.

1
Retrieve
Hybrid vector + BM25 search with cross-encoder reranking, tuned on a fixed eval set rather than by feel.
2
Evaluate
LLM-as-a-judge scoring for faithfulness and citation coverage, versioned alongside the prompts it grades.
3
Ship
GitHub Actions gates on those scores, so a regression blocks the merge. The demo is live, not a screenshot.

3. Figures (Projects)

Seven projects, numbered as figures. Each links to the artifact it describes — a live app, a repo, or a paper — rather than a screenshot of one.

Fig. 1 · Production RAG

AskMyDocs: Production-Grade RAG with CI-Gated Evaluation

Fig. 1 — AskMyDocs retrieval and CI-gate architecture. Modules (loader, chunker, embedder, vectorstore, reranker, LLM, prompts) are independently swappable.

Ask questions of your own PDFs and get answers with inline citations. Retrieval is hybrid (60% ChromaDB vector similarity, 40% BM25) with the top 10 candidates re-scored to the top 3 by an ms-marco-MiniLM-L-6-v2 cross-encoder. Any answer that comes back without citations is flagged as a hallucination, and prompt templates are versioned (default / strict / concise) and selectable per query. Every release is graded by an LLM judge in GitHub Actions: faithfulness under 0.70, relevance under 0.70 or citation rate under 0.80 fails the build and the version never ships. The whole pipeline is built as swappable modules (loader, chunker, embedder, vectorstore, reranker, LLM, prompts) so any stage can be ablated on its own. Live and public on Hugging Face Spaces.

Table 1 — CI gate thresholdsverifiable in repo
Metric	Gate
Retrieval mixvector similarity / BM25	60 / 40
Faithfulnessbuild fails below this	≥ 0.70
Relevancebuild fails below this	≥ 0.70
Citation ratebuild fails below this	≥ 0.80

LLaMA (Groq) · ChromaDB · BM25 · Cross-Encoder · GitHub Actions · Gradio

Live demo ↗Source ↗

Fig. 2 · Computer Vision · Robotics

Vision-Based Virtual Maze Navigator

An e-puck robot solves an unmapped Webots maze from camera pixels only: no odometry, no map, no distance sensors in the control loop. The finding was counter-intuitive: wall proximity registers as edge density collapsing toward zero rather than spiking, and the detector built on that survived a blur test unmodified. Per-frame CSV telemetry surfaced six failure modes; five are fixed on main, and the last traced to the one detector keyed on absolute intensity rather than relative structure, with a relative-brightness fix specified that holds through a 50% illumination drop.

Illustrative — detector confidence vs. corridor illumination, not plotted telemetry.

Now the seed project for faculty-advised research targeting IEEE RA-L. See revision history →

OpenCV · Webots (e-puck) · Canny Edges · HSV Detection · Reactive Control · NumPy

Demo video ↗Write-up ↗Source ↗

Fig. 3 · LLM Research · Publication

LLM-Generated Social Networks: Cross-Cultural Study (Capstone)

Replicated and extended an ICWSM 2025 paper on whether LLMs generate structurally realistic social networks. Built a full generation-analysis-benchmark pipeline across a 4×4×4×3 matrix (prompting × cultures × languages × GPT-4.1 tiers, two seeds per condition) producing 192 verified directed networks from 50 demographically grounded personas. Political affiliation dominates tie formation under three of four methods, but the global method substitutes age, which means homophily results reported under a single prompting scheme may be design artifacts. First-author paper, under review at NeurIPS 2026.

GPT-4.1 · NetworkX · pandas · Experiment Design

arXiv:2605.12898 ↗

Fig. 4 · Deep Learning

RUL Prediction with LSTM: Uncertainty-Aware Predictive Maintenance

Benchmarked four recurrent architectures (Vanilla LSTM, Stacked BiLSTM, LSTM-Attention, CNN-LSTM) on NASA C-MAPSS for turbofan Remaining Useful Life. Monte Carlo Dropout and deep ensembles supply the uncertainty, and maintenance decisions are taken on a mean − 2σ lower bound rather than a point estimate. Gradient-based attribution and temporal attention heatmaps give per-prediction explanations that can be audited, and zero-shot transfer was evaluated across C-MAPSS subsets.

PyTorch · TensorFlow · Uncertainty · XAI

Source ↗

Fig. 5–7 · Supporting work

5. Commonsense Reasoning with Pre-trained Language Models

Benchmarked RoBERTa-MNLI and OPT-1.3B on PIQA and related commonsense tasks, zero-shot vs. fine-tuned, with Hugging Face Transformers.

RoBERTa · OPT-1.3B · PyTorch · Hugging Face

Source ↗

6. Hate Speech Detection

Text classification pipeline with Logistic Regression and Random Forest, reaching 92% accuracy on a Kaggle-sourced dataset for content-moderation use cases.

Python · NLP · scikit-learn

Source ↗

7. Abstractive Text Summarizer

Streamlit app condensing long articles with a distilBART (sshleifer/distilbart-cnn-12-6) seq2seq model, summarizer logic factored out for reuse.

Hugging Face · distilBART · Streamlit

Source ↗

Plus smaller ML and annotation work: spam detection, sentiment annotation, submarine life simulation, classic-algorithm notebooks. Browse all repositories ↗

4. Revision History (Work Experience)

Work experience, kept as a paper keeps its versions: what changed, and why it needed a new one.

v2Research Collaboratorcurrent

Aug 2026 – Present

University of Arizona

Advisor: Prof. Eungjoo Lee, School of Electrical, Computing & Software Engineering · Target venue: IEEE RA-L

Investigating whether a vision-only mobile robot can recognize when its visual evidence is unreliable for navigation and adapt its behavior before an unsafe or incorrect decision, generalizing a failure mode found in the Vision-Based Virtual Maze Navigator.
Built a multi-environment Webots evaluation benchmark spanning 16 maze layouts (11 training, 5 held-out test) under 8 visual corruption types (low illumination, motion blur, defocus blur, occlusion, noise, brightness/contrast, reduced field of view, plus clean) at 5 severity levels each, informed by the RobustNav (ICCV 2021) evaluation protocol.
Built an automated data-logging pipeline capturing per-frame camera imagery, perception outputs, actions, pose and collision/success outcomes, with simulator ground truth reserved for training labels and evaluation rather than exposed to the navigation policy at test time.
Ran the baseline Canny/HSV controller across 792 trials and uncovered a structural wall-following loop-trap masked by single-maze testing, plus sharply differential corruption sensitivity (4–8% collision under blur vs. 60–84% under illumination, brightness/contrast, occlusion and noise) that's now driving a controller redesign and a relative-threshold baseline comparison.

Webots · OpenCV · Robustness Benchmarking · RobustNav ProtocolCode ↗

v1Data Analyst Intern

Jan 2025 – Aug 2025

Sudhir Mehrotra & Associates, Chartered Accountants · Bareilly, India (Hybrid)

First hands-on exposure to production data pipelines: the itch that led into the ML/LLM research and engineering work featured throughout this site.
Built Python ETL pipelines that automated financial workflows, cutting manual processing effort by 30%.
Developed a time-series cash-flow forecasting module that improved estimation accuracy by 15% over baseline.
Automated recurring Excel reporting with Python and VBA, eliminating multi-hour weekly manual tasks for the audit team.
Designed structured data-reporting systems for audit and compliance teams, improving traceability across reviews.

Python · ETL · Time-Series Forecasting · Excel / VBA

5. Results (Skills)

Tools and methods, grouped the way they get used together rather than alphabetized.

Category	Tools & methods
LLMs & RAG	OpenAI API · GPT-4.1 · LLaMA (Groq) · RAG pipelines · Hybrid search · Cross-encoder rerank · ChromaDB · sentence-transformers · Prompt engineering · Citation grounding
LLM Eval & Reliability	LLM-as-a-judge · Hallucination detection · Faithfulness / relevance · Citation metrics · CI-gated thresholds · Versioned prompts
Deep Learning	PyTorch · TensorFlow · Keras · scikit-learn · HF Transformers · LSTM · Attention · MC Dropout · Deep ensembles · Uncertainty · XAI
Computer Vision & Robotics	OpenCV · Canny edge detection · HSV color detection · Webots (e-puck) · Reactive control policies · Robustness benchmarking · Visual perturbation testing
Languages	Python · SQL · R · Bash · JavaScript
Data & Viz	pandas · NumPy · NetworkX · Matplotlib · Seaborn · Jupyter · ipywidgets · Gradio · Streamlit
MLOps & Tools	Git · GitHub Actions (CI/CD) · AWS · Linux · HF Spaces · REST APIs · RAPIDS (GPU) · LaTeX
Databases	MySQL · PostgreSQL · MongoDB · ChromaDB (vector)

6. Appendix (Education & Certifications)

A. Education

M.S. in Data Science

Apr 2024 – May 2026

University of Arizona · Tucson, AZ

GPA: 3.78 / 4.00 · Conferred May 2026.

Machine Learning · Applied Natural Language Processing · Computational Linguistics · Data Mining & Discovery · Foundations of Data Science · Data Analysis & Visualization · Data Ethics

B.Tech, Computer Science & Engineering (Software Engineering)

2020 – 2024

SRM Institute of Science and Technology · Chennai, India

CGPA: 8.51 / 10.00.

Data Structures & Algorithms · Design & Analysis of Algorithms · Artificial Intelligence · Compiler Design · Operating Systems · Database Management Systems · Formal Languages & Automata · Probability & Queueing Theory · Discrete Mathematics · Design of Experiments · Calculus & Linear Algebra

B. Certificates

AI Fluency: Framework & Foundations

Anthropic · Jun 2026

4D framework for working with AI · Delegation & task decomposition · Description (prompting/context/iterative refinement) · Discernment & diligence

Verify ↗

Claude 101

Anthropic · Jun 2026

Claude model family & capabilities · Effective prompting/context management · Projects/artifacts/extended workflows · Practical/responsible use

Verify ↗

Fundamentals of Accelerated Data Science

NVIDIA · Oct 2025

GPU-accelerated computing (RAPIDS) · Core data science · Applied ML · Accelerated ecosystem integration

Verify ↗

AWS Academy Cloud Operations

Amazon Web Services · Nov 2022

Cloud infrastructure operations · Monitoring & management on AWS · Deployment & automation fundamentals

Verify ↗

7. References (Publications)

Under review · NeurIPS 2026First author

When Do LLMs Generate Realistic Social Networks? A Multi-Dimensional Study of Culture, Language, Scale, and Method

R. Upadhyay, S. H. Kilaru, S. T. Manikyala, S. S. K. Ramavath, S. Nunavathu, D. Alharthi

May 2026 · arXiv:2605.12898 [cs.SI]on arXiv

Building on homophily and structural balance theory, we formalize four LLM-based tie-formation mechanisms (sequential, global, local, and iterative) as distinct conditional distributions over edge sets. Using a fixed roster of 50 demographically grounded personas, we generate 192 verified directed networks across four cultural contexts, four prompt languages, three GPT-4.1 variants and four prompting architectures, with two seeds per condition.

F1.Cultural framing measurably shifts inbreeding homophily and largest-component connectivity.
F2.Political affiliation dominates tie formation under three of four methods, while the global method substitutes age. Prompt architecture is therefore a substantive sociological variable: homophily findings reported under a single prompting scheme may be design artifacts.
F3.Model scale produces a divergence ranking that reproduced across all three studies (GPT-4.1 ↔ mini = 0.074, ↔ nano = 0.119), with the smallest variant differing qualitatively rather than only in noise.

GPT-4.1 · NetworkX · pandas · Experimental Design

Abstract ↗PDF ↗

8. Correspondence (Contact)

I finished my M.S. in May 2026 and I'm looking for AI / ML engineering and research roles in the US, while working as a Research Collaborator at the University of Arizona. If you're hiring, or you want to talk about LLM evaluation or reliability under perturbation, email is the fastest way to reach me.

Emailraghav0408upadhyay@gmail.com

LinkedInraghavupadhyay04

GitHubraghav-upadhyay2002

Hugging Faceraghavupadhyay

arXivupadhyay_r_1