> init --user
I build production-grade LLM systems, agentic infrastructure, and real-time AI that drive measurable impact on automation and user engagement.
Riverside, CA
> ls ./projects/ — recent shipped work
Production-grade multi-agent financial advisor powered by LangGraph, Claude, ARIMA forecasting, Markowitz optimisation, and Isolation Forest anomaly detection. Deployed on Hugging Face Spaces with a FastAPI backend and a custom chat UI.
Open-source RAG library that applies particle physics significance testing to eliminate hallucinations. Estimates the background noise distribution in embedding space and only retrieves chunks that clear a configurable σ threshold.
Drop-in sktime forecaster that uses a ReAct agent loop to automatically select and configure time series pipelines from natural language descriptions. Supports Claude, GPT-4o, and Gemini backends.
> cat ./posts/* — notes from building real AI systems
Every line of code, every reason behind it. Attention is a routing mechanism — by the end of an attention pass, a token's vector is a weighted mixture pulled toward whatever it attended to.
The heart of every modern LLM. Before attention, RNNs processed sequences one token at a time — a rolling summary that forgot the distant past. Attention fixes that.
Token IDs are just integers — meaningless to a neural network. Embeddings map them to dense vectors where relationships are encoded in direction and distance.
> more posts shipping soon — say hi via Get in Touch below.
> open ./deep-dives/ — ML & AI fundamentals, explained properly
> topic_01
From the original paper to multi-head attention, positional encoding, and why the architecture took over everything.
> topic_02
How retrieval-augmented generation works, what breaks in production, and how to design a retrieval pipeline that holds up.
> topic_03
A practical guide to evaluating LLM outputs — metrics, frameworks, LLM-as-judge, and building eval suites that catch real regressions.
> topic_04
Decision Trees, Random Forests, XGBoost, SVMs, logistic regression, regularization, bias-variance, cross-validation, and when classical beats deep learning.
> topic_05
The foundations that everything in ML is built on — distributions, Bayes, hypothesis testing, and how they connect to real modelling decisions.
> topic_06
Word2Vec, GloVe, Sentence Transformers, contrastive learning, embedding geometry, similarity metrics, PCA, and UMAP — with worked exercises.
> git log --all --oneline
> Jul 2024 — May 2026
Engineered an agentic RAG system (LangChain + Postgres pgvector) for natural-language → SQL analytics. Built an LLM-powered QA framework reaching 80% automation, a virtual persona simulation framework, and architected a multi-agent commerce system on Vertex AI + LangGraph.
> Jan 2024 — Jul 2024
Shipped an LSTM-based sequential prediction model for long-term content performance. Led a comment ranking system using semantic embeddings + LLM summaries that lifted CTR by 4.5% and comment volume by 4.0%.
> Dec 2023 — Feb 2024
Automated reliability assessment of medical research papers using LLM-based algorithms. Improved GPT-4 evaluation accuracy to 75% via prompt optimization and pushed Mixtral to 69% with QLoRA fine-tuning.
> Jun 2023 — Sep 2023
Trained a self-supervised audio embedding model on 35K+ hours of audio (86% downstream accuracy). Engineered a multimodal embedding space aligning audio and text, beating text-only baselines by 5 points.
> cat ./about.md
I’m Kuntal — an AI engineer working on LLM systems, retrieval pipelines, and multi-agent infrastructure. I like turning messy, ambiguous problems into clean, production-shaped products.
Until recently (May 2026), I was at Beehive AI, where I built agentic RAG systems, LLM-powered evaluation frameworks, and virtual persona simulators that help product teams make sharper decisions.
Before that, I spent seven years doing theoretical and computational particle physics for my Ph.D. at UC Riverside, with valuable research experiences at Deepgram (multimodal audio embeddings), PI School in Rome (LLM evaluation for medical research), and Pocket FM (sequential prediction + comment ranking).
University of California, Riverside
Ph.D., Physics
Sep 2017 — Jun 2024
Indian Institute of Science Education and Research, Kolkata
BS-MS Dual Degree, Physics
Aug 2012 — May 2017
> ls -la ./stack/ — tools I reach for most often
> open inbox — reply latency: low
My inbox is always open — whether you have a project in mind, a research collaboration, or just want to say hi, I’ll do my best to reply.
Say Hello> kuntal.beehiveai@gmail.com · +1 (951) 202-8635