LLM fine-tuning
Models shaped for your task.
Fine-tune and adapt open language models for your domain, tone or workflow. From data preparation and training to evaluation and local model formats.
INDEPENDENT AI ENGINEERING & DESIGN STUDIOBASED IN INDIA. BUILDING EVERYWHERE.
AI engineering. Thoughtful design. Zero guesswork.
Fine-tuned models. Grounded AI systems.
Thoughtful digital experiences.
From the first experiment to the final product.
FROM MODEL TO EXPERIENCE CONNECTED BY DESIGN
THE ZEROSHOT SYSTEM
01 / MODELS & DATA
Model pipelineEVALUATED
02 / THE INTELLIGENCE
✳ZeroShot AI
“What does the evidence say?”
↳ Retrieve. Verify. Respond.
● ● ●03 / THE EXPERIENCE↗
your product.Model Interface ↗
Ideas into
interfaces.
interfaces.
SCROLL TO DISCOVER
AI engineering at the core.
Design in every detail.
01 / THE IDEALESS FRICTION. MORE POSSIBILITY.
From model weights
to the last pixel.
Built with intent.
ZeroShotLab is an independent AI engineering, web design and Android development studio led by Randhir Kumar in India. I build LLM fine-tuning workflows, retrieval-augmented generation (RAG) systems, AI agents, web experiences and Android apps.
My published work includes HumanFlow, a fine-tuned Llama 3 8B model; Corrective RAG, a document question-answering system with passage verification; and MoEVL-Tiny, a compact vision-language model. Explore the code, model weights and research below.
What we can build together02 / AI ENGINEERING IN THE OPENMODELS. SYSTEMS. SOURCE CODE.
Beyond the interface.
Into the intelligence.
Selected engineering work, published with source code or model weights. Explore how it works, not just how it looks.
01 / LLM FINE-TUNING
HumanFlow
A fine-tuned Llama 3 8B model that rewrites stiff AI drafts into natural prose. Published with GGUF versions for local use.
02 / RETRIEVAL & VERIFICATION
Corrective RAG
Document Q&A that checks retrieved passages, cites its sources and declines when the evidence is missing. Built with FastAPI, Docker and tests.
03 / VISION-LANGUAGE MODELS
MoEVL-Tiny 0.5B
A compact vision-language model pairing SigLIP with Qwen2.5-0.5B and mixture-of-experts layers. Includes a Gradio demo Space.
03 / WEB EXPERIENCESDESIGNED TO BE EXPERIENCED.
A little less talk.
A lot more tangible.
Working concepts for real-world businesses. Open a demo. Explore the details. Meet the assistant.
Concept work, not client commissions. Businesses are not affiliated with ZeroShotLab.
04 / AREAS OF PRACTICEDESIGN × ENGINEERING × INTELLIGENCE
Looks considered.
Works harder.
One engineer. From the first sketch to the last detail. Everything connected, everything yours.
Tell me what you’re buildingRAG & knowledge systems
Answers with something behind them.
Build retrieval and document question-answering systems grounded in your own information, with source citations and checks for weak evidence.
AI agents & automation
AI that fits the way you work.
Build assistants and tool-connected workflows that help with real tasks. Connect your knowledge, APIs and existing tools, with clear boundaries and human handoffs.
Web design & development
The engineering. Made approachable.
Design and build websites and interfaces that make your product clear, useful and memorable. Bring the model and the experience together.
Android app development
Your product. In their pocket.
Build Android apps with thoughtful interfaces, API integrations and AI features where they help. From the first prototype to testing and release preparation.
Different challenges.
A conversation first.
Tell me what you want to build, what you’ve tried and where you’re stuck. We’ll work out the right approach and scope together.
Let’s talk about your projectTHE PROCESSNO BLACK BOXES.
A clear path.
From hello to live.
Start with a conversation. See the direction before you commit. Stay close to the work at every step.
- 01
See the possibility.
Talk through the problem, the users and the outcome. Review the data, existing system or early idea together.
- 02
Agree on the details.
Define the scope, milestones, timeline and what a useful result looks like.
- 03
Build, together.
Review prototypes, model outputs or working previews as the project takes shape. Test against the goals we agreed on.
- 04
Make it yours.
Your code, your accounts and a clear handover. Documentation and next steps agreed as part of the project.
05 / INDEPENDENT BY DESIGNA small studio.
A small studio.
A direct line.
I’m Randhir Kumar, an AI engineer from India. I fine-tune and deploy language models, build agents and retrieval systems, and I’m comfortable down at the hardware level with IoT work. I’ve also shipped five Android games end to end, which taught me to care about the last 10% that users actually feel.
ZeroShotLab is my studio, registered in India as an MSME (Udyam). The name is a nod to zero-shot learning: getting useful results without endless setup. The panda is just me. I like pandas.
I use AI tools to speed up routine work and review everything I deliver myself. If your project restricts AI tools, tell me.
More from the lab All projects & experiments ↗
06 / OPEN SOURCE
Selected work, in the open
Code and models I built and published on GitHub and Hugging Face. Each link opens the source, so you can read exactly how it was built.
01Python · NLI · FastAPI · Docker
02FastAPI · Streamlit · Groq LLMs
03Python · Streamlit · NLP
04Llama 3 8B fine-tune · Hugging Face
05Python · RAG evaluation · paper
06Node.js · Express · Socket.io
07C++ · ESP32 · round TFT display
08PyTorch · MoE · Hugging Face
Corrective RAG
A question-answering system over your own documents that checks each retrieved passage before answering, cites its sources, and declines to answer when the evidence is not there. Built for small open models, with a FastAPI server, Docker setup and tests.
Nova, a landing-page chatbot
A website chatbot that greets visitors, answers product questions and points them toward signing up. FastAPI backend, Streamlit chat interface, with a live demo.
AI resume analyzer
Upload a PDF resume and paste a job description to get a keyword and skill match, section and formatting checks, and a semantic match score. Runs as a Streamlit web app with a live demo.
HumanFlow
A fine-tuned open language model that rewrites stiff AI drafts into more natural, readable prose. Published on Hugging Face with quantized GGUF versions for running locally.
Verification without sufficiency
Research on why checking passages one by one breaks down on multi-step questions in RAG systems, and how splitting the question into parts fixes it. Includes code, results and a written paper.
Real-time device tracking
A web app that shows connected devices moving on a live map, updated in real time over websockets.
Golubot desk-pet firmware
Firmware for a small ESP32 desk robot with animated eyes, moods, a Pomodoro timer and mini games, driven by a touch sensor.
MoEVL-Tiny 0.5B
A small vision-language model that describes and answers questions about images, built to be light enough for on-device use. It pairs a SigLIP image encoder with Qwen2.5-0.5B and mixture-of-experts layers, and has a Gradio demo Space.
See everything on GitHub (@iamhero2709) → Models on Hugging Face (randhir302) →
08 / AFTER HOURS
Five Android games, in testing
I designed and built five games solo. They’re in closed/internal testing on Google Play now; links go up here once they’re public.
#1
Coming soon to Google Play
#2
Coming soon to Google Play
#3
Coming soon to Google Play
#4
Coming soon to Google Play
#5
Coming soon to Google Play
06 / RESEARCH & WRITING
Research preprints
Papers I wrote on retrieval, verification and reasoning, posted on arXiv.
01Sahil Kumar Rajput, Randhir Kumar · Oct 2026 · arXiv:2610.27366
02Randhir Kumar · Aug 2026 · arXiv:2608.00585
03Randhir Kumar · Jun 2026 · arXiv:2606.24948
Trail-Weighted Influence Scores for Approximate Outcome Prediction in Bayesian Networks
A way to speed up predictions in Bayesian networks by dropping evidence that barely affects the outcome, with error bounds and experiments.
Verification Without Sufficiency: Per-Chunk Filtering Fails on Multi-Hop RAG, and Decomposition Repairs It
Shows why checking retrieved passages one at a time fails on multi-step questions in RAG systems, and that breaking the question into parts fixes it.
Holographic Memory for Zero-Shot Compositional Reasoning in Knowledge Graphs: A Mechanistic Study of Where and Why It Fails
Tests whether holographic memory can answer unseen multi-step knowledge-graph questions, and explains why it does not.
07 / BEFORE WE BEGINA few things
A few things
worth knowing.
What does ZeroShotLab do?
ZeroShotLab is Randhir Kumar’s independent AI engineering, web design and Android development studio in India. Work includes LLM fine-tuning, retrieval-augmented generation (RAG), AI agents, model evaluation, web interfaces and Android app development. Projects start with a conversation about your goal and current setup.
Should my project use RAG or LLM fine-tuning?
RAG retrieves relevant information at answer time and is useful when answers need current or private documents and source references. Fine-tuning changes a model’s behavior using training examples and can help with a specific task, style or output format. A project may use either or both; the choice depends on the data and evaluation goals.
What AI engineering work can I review?
HumanFlow is a fine-tuned Llama 3 8B model published on Hugging Face. Corrective RAG is an open-source document Q&A system with passage verification and source citations. MoEVL-Tiny 0.5B is a compact vision-language model. The engineering portfolio links directly to the code or model pages.
How do you evaluate an AI system?
Evaluation should reflect the task. For a retrieval system, check whether it finds the right evidence, cites it correctly and declines unsupported answers. For a fine-tuned model, compare it with the base model on representative held-out examples. We agree on the evaluation criteria as part of the project scope.
Can you work with my existing models or codebase?
Yes. Share your current model, data sources, codebase or API setup and what you want to improve. I can help with a focused part of an existing system or discuss a new implementation based on the scope.
Who owns the code and project accounts?
You retain your project code and accounts as agreed in the project scope. Third-party models, datasets and dependencies keep their own licenses, which need to be checked for the intended use.
How do we start a project?
Use the project enquiry form or email randhir2709vns@gmail.com. Share your goal, relevant links, current stack and constraints. We discuss the technical approach, scope and timeline before agreeing on the work.
HAVE SOMETHING IN MIND?
Let’s build
your next move.
Have a model to adapt, a retrieval system to build, or a digital experience in mind? Tell me what you’re working on. Let’s discuss the possibilities. Or email randhir2709vns@gmail.com.