Open to work, 2026
Curious by default.
Shipping by habit.
I'm Rohit Raju, an AI Engineer and ML Researcher shipping agentic LLM systems meant for production. At Per Diem (YC W21) I deployed a Bedrock-based multi-agent platform across 500+ retail stores, with 30% weekly order growth and up to 70% retention.
Projects I'm proud of.
A tight selection across LLMs, statistical modeling, and AI for education. Each one shipped, with paper, code, or video.
-
Improving the Generalizability of Models of Collaborative Discourse
Methods to make LLM classifiers of student talk generalize across grade levels, demographics, and curricula. Embedding-augmented RoBERTa fine-tuning and Mistral-embedding SVMs significantly outperformed vanilla fine-tuning across five diverse datasets. Presented at EDM 2025, Palermo.
-
TaleTutor, classroom PDFs as narrative learning
A GPT + LangChain RAG system that turns classroom PDFs into narrative-driven, zero-shot responses, so learning feels like a story, not a search. Finalist at the UC Berkeley AI Hackathon.
-
DataFlow Architect, Data Science as a Service
An AI-driven platform that automates dataset profiling, EDA, and ML use-case synthesis using GPT-4o. A prompt-driven, dockerized reporting system with latency-cost benchmarking across LLMs and on-demand PDF export, cutting manual workflow design time by 60%.
-
YouTube titles & views, a statistical approach
Hypothesis testing on whether title phrasing measurably shifts view counts. Full report, code, and a recorded presentation walking through the findings.
-
Auto Semantic & Syntactic Grader
An evaluator that grades student code on both semantic correctness and syntactic structure, cutting TA grading time while keeping feedback meaningful.
Peer-reviewed publications.
Work at the intersection of language models, error correction, and collaborative learning.
-
IEEE I2CT2023
A System for Enhancing Accuracy of Noisy Text Using Deep Network Language Models
BART & MarianMT applied to OCR outputs, 35% reduction in word-error rate.
Read paper ↗ -
IEEE I2CT2024
Comparative Study on Synthetic and Natural Error Analysis with BART & MarianMT
Identifies a 26% break-even point where synthetic errors begin to mislead evaluation.
Read paper ↗ -
JIKM (World Scientific)2024
Grammatical vs. Spelling Error Correction: Responsiveness of Transformer Language Models
BART outperforms MarianMT on spelling correction by 24.6%; behavior diverges on grammar.
Read paper ↗ -
EDM2025
Improving the Generalizability of Models of Collaborative Discourse
Cross-dataset generalization for classifiers of student talk in collaborative learning.
Read paper ↗ -
ISLS2025 · Accepted
Facilitating Productive Uncertainty in Small-Group Jigsaw Activities
How AI-driven feedback can foster productive uncertainty in small-group learning.
Accepted
Why I do this.
I'm an AI Engineer and ML Researcher focused on agentic LLM systems. Lately that means real merchant workflows: support, analytics automation, and campaign optimization. The hard part is reliability and scale, not a slick demo.
At Per Diem (YC W21), I designed and deployed a production-grade multi-agent platform on Amazon Bedrock, built on agent orchestration, RAG, and NL-to-SQL, now used by 500+ retail stores. It helped drive 30% weekly order growth and up to 70% retention across stores running on it.
Through the NSF AI Institute for Student-AI Teaming at CU Boulder, I work on NLP and collaborative learning systems. Five AI papers total: two with the institute, plus three first-author papers from undergrad spanning transformer-based models, interpretability, and applied AI in education.
What I'm actually chasing is AI for education. Not chatbot tutors, but personalized narrative, learning that adapts to the student in front of it and eventually lets them rewrite their own story. That's the long bet.
Experience
-
Jul 2025 - Apr 2026
AI Engineer
Per Diem (YC W21) · New York, USA
Architected a production-grade multi-agent AI system on Amazon Bedrock with serverless infrastructure, safety boundaries, and secure deployment pipelines, now live across 450+ retail stores, driving 30% weekly order growth and up to 70% retention through AI-powered campaign optimization.
Built the agent orchestration layer (Bedrock Agents + few-shot prompting) to route natural-language queries across Support, Analytics, and Marketing agents. Implemented a RAG-powered support workflow on Pinecone + Bedrock Knowledge Bases. Designed an NL-to-SQL analytics pipeline via tool calling (schema retrieval + read-only execution). Shipped an autonomous Marketing AI agent that recommends and deploys margin-aware, time-optimized campaigns.
-
Oct 2023 - Apr 2025
Machine Learning Research Assistant
NSF AI Institute for Student-AI Teaming · Institute of Cognitive Science, CU Boulder
Built classification models that detect collaborative-discourse patterns in student conversations, enabling real-time feedback in K-12 classrooms. Trained and evaluated 6 Mistral-based variants (Few-shot, LoRA, Mistral+SVM) across four diverse datasets spanning human-coded and ASR-transcribed data, 14% average AUROC improvement in cross-domain generalization over a RoBERTa baseline.
Integrated LIME-based interpretability into evaluation pipelines to analyze token-level feature contributions. Contributed to an AI moderation agent for K-12 classrooms using zero-shot prompting on AWS LLaMA 70B with safety guardrails.
-
Nov 2022 - Aug 2023
PLC Simulations Systems Designer (Intern)
BOSCH Research
Designed and implemented the SimuBridge backend in C# to emulate PLCs and feed clean data into DeviceBridge. Integrated with the OPC server, accelerating DeviceBridge stress testing by 50%.
Education MS Data Science, University of Colorado Boulder, GPA 4.0/4.0 (2023-2025) B.Tech Computer Science, Amrita School of Computing, GPA 8.68/10 (2019-2023)
What people say.
"Rohit served as a teaching assistant for my Neural Networks and Deep Learning course, preparing a hands-on TensorFlow session, managing research volunteering for a large student group, and mentoring two project teams. He's confident, self-disciplined, and a committed leader with strong time management and communication skills."
"Henri Frederic Amiel could have had Rohit in mind when he said, 'Doing easily what others find difficult is talent; doing with talent what is impossible is genius.' Amrita School of Engineering feels blessed to nurture talented students like him."
Off the clock, as FROST tube.
Outside of work I run a vlog channel about everyday life, travel, and the odd small story. Different craft from the day job, but the same itch to make things.
Let's build something.
Research collaborations, AI engineering, or just a conversation about where this is all going. My inbox is open.
rohit.raju@colorado.edu