An applied AI & data science studio · Est. 2026

We build the AI systems
that actually ship. — Not slide decks. Production code.

Kratvanta is a small specialist team. We build generative AI, computer vision, and forecasting systems — and advise companies on what to build, why, and when. IEEE-published research, ex-ISRO, ex-PetPooja, ex-OCTO Advisory.

Team
3 engineers
Discipline
Build · Ship · Advise
Published research
IEEE · LLM evaluation
Based in
Gujarat, India
AI feasibility LLM fine-tuning YOLOv8 detection ML roadmaps OCR pipelines Demand forecasting RAG systems Scoping sprints Power BI dashboards Time-series modelling Computer vision Data audits AI feasibility LLM fine-tuning YOLOv8 detection ML roadmaps OCR pipelines Demand forecasting RAG systems Scoping sprints Power BI dashboards Time-series modelling Computer vision Data audits

Most AI consulting is a 60-page deck a pilot that never shipped. We do the opposite — small team, deep technical work, production systems that earn back their cost in months, not years.

Services

Four practices, done deeply.

We deliberately don't do everything. These are the disciplines where our team has shipped real production work — and where we can defend our recommendations technically, not just rhetorically.

01

Generative AI &
LLM Engineering

Custom chatbots, RAG systems, document intelligence, and fine-tuned LLMs for your domain. We've shipped LLaMA-based chatbots in production and published IEEE research on Llama-2 vs Mistral-7B fine-tuning with LoRA & PEFT.
LLaMA Mistral RAG LoRA / PEFT Prompt eng. Gemini API
02

Computer Vision
& ML Systems

Object detection, classification, OCR pipelines, and end-to-end ML systems. We've built YOLOv8 inventory monitors, OCR systems trained on 100,000+ images at 93% accuracy, and CNN-based classifiers deployed in production.
YOLOv8 PyTorch CNN / RNN OpenCV Transfer learning Whisper
03

Forecasting &
Data Analytics

Demand forecasting, time-series models, executive dashboards, and data pipelines. From 6M+ row traffic datasets to Power BI dashboards driving operational decisions — we turn messy data into systems decision-makers actually use.
Bi-LSTM Q-learning Power BI SAS Viya Time-series Data pipelines
04

AI Advisory
& Strategy

For organizations that need clarity before code. We audit your data, assess feasibility of proposed AI initiatives, and produce concrete roadmaps with realistic timelines, costs, and failure modes. Built for companies who want a partner that will tell them when ML isn't the right answer.
Feasibility audits AI roadmaps Data audits Scoping sprints ROI modelling Workshops
Selected work

Things we've actually built.

Production systems shipped during our previous engagements at foodtech, urban-planning, research, and analytics firms. Real numbers, real outcomes.

Foodtech · GenAI + Computer Vision Flagship

AI-Powered Smart Kitchen &
Inventory Optimization

Built an end-to-end system to fight overstocking and food spoilage at a multi-outlet kitchen operation. Deep learning models forecast daily sales and waste; a Q-learning agent integrated with YOLOv8 and Gemini API handles automated reordering, ingredient detection, and real-time spoilage monitoring via camera feeds.

20%
Food waste reduction
YOLOv8
Spoilage detection
RL
Auto-reordering agent
Python PyTorch YOLOv8 Q-learning Gemini API Flask
Research · LLM Eval IEEE Published

VITCHAT — Llama-2 vs Mistral-7B Comparative Study

A comparative evaluation of Llama-2-7b and Mistral-7b for domain-specific conversational AI, with LoRA and PEFT applied for parameter-efficient fine-tuning on consumer GPUs. Findings presented and published at IEEE.

0.43
BLEU score
IEEE
Published
Llama-2-7b Mistral-7b LoRA PEFT Transformers
Urban analytics · Forecasting Production

Traffic Forecasting &
Route Optimization

Analyzed 6M+ traffic data points to identify congestion patterns and forecast bottlenecks. Outputs delivered to stakeholders via a Power BI dashboard translating model predictions into operational decisions.

93%
Forecast accuracy
6M+
Data points
Python Time-series Power BI
Industrial OCR · Document AI 100k+ images

OCR Pipeline +
GenAI Recipe Automation

Replaced a traditional OCR-only workflow with a hybrid system combining custom-trained PyTorch models with Generative AI for recipe generation. The OCR system was optimized on a corpus of 100,000+ images; LLM integration moved the workflow from extraction to generation.

93%
OCR accuracy
100k+
Training images
GenAI
Recipe generation
PyTorch OCR LLMs Custom CNN
Agritech · Computer Vision Field deployment

Plant Disease Detection
at Leaf-Level

A CNN-based classifier paired with YOLO-based localization for plant species identification and disease prediction. The system highlights affected regions with bounding boxes so farmers can act on specific leaves — not just receive a label.

80%+
Classification accuracy
YOLO
Lesion localization
TensorFlow YOLO Flask CNN
Research · ISRO Space tech

Root-Zone Soil Moisture
Estimation for ISRO

A machine learning tool estimating root-zone soil moisture from satellite-derived surface soil moisture data. Built during a research engagement with the Indian Space Research Organisation.

ISRO
Research engagement
8 mo.
Duration
Python Scikit-learn Remote sensing
The team

Three people. No middle layer.

When you hire Kratvanta, you talk to the people writing the code. No account managers, no offshore handoffs, no junior team you've never met doing the actual work.

01

Hiren Thakkar

Data Scientist · AI Engineer

Data scientist with a published IEEE research background. Specializes in scaling generative AI, NLP, and predictive modeling to production. Has shipped LLaMA-based chatbots, OCR pipelines on 100k+ images, and LSTM-based sentiment models.

MSc Data Science, VIT (CGPA 7.93)
IEEE-published research on LLM fine-tuning
Ex-OCTO Advisory, ex-Prayosha, ex-Isnartech
LLMs NLP PyTorch AWS SageMaker
LinkedIn → Portfolio →
02

Swapnil Sheth

Data Analyst

Data analyst with strong technical breadth — builds executive dashboards on 40k+ row datasets, designs forecasting models that drive operational decisions, and has contributed YOLOv8 detection and CNN classifier work on production projects. The bridge between raw data and decisions stakeholders can act on.

MSc Data Science, DAU Gandhinagar
Co-built the Smart Kitchen system at Prayosha
15-20% prediction accuracy lift across projects
Power BI Forecasting Data pipelines ML breadth
LinkedIn →
03

Vijita Dudhrejiya

AI Advisory & Solutions

Research-grounded AI advisor with cross-domain experience — space tech, foodtech, fintech-adjacent analytics. Helps clients move from "we should use AI somewhere" to a concrete, technically-honest roadmap. Strong opinions on when ML earns its place and when it doesn't.

MSc Data Science, DA-IICT (CGPA 8.0)
Ex-ISRO research, ex-PetPooja, ex-Profond AI
SAS-certified · Machine Learning Specialist
AI Strategy Solution Design Research-led Cross-domain
LinkedIn → Email →
Technical stack

The tools we actually use.

No buzzword soup. Every item below is something one of us has shipped to production. If your stack isn't here, ask — we may still be a fit.

Languages

  • Python
  • R
  • SQL
  • SAS
  • PySpark

ML / DL

  • PyTorch
  • TensorFlow / Keras
  • Scikit-learn
  • Hugging Face
  • SAS Viya 3.5

GenAI / LLM

  • LLaMA-2 / LLaMA-70B
  • Mistral-7B
  • Gemini API
  • LoRA / PEFT
  • Whisper (ASR)

Computer Vision

  • YOLOv8
  • OpenCV
  • Custom CNNs
  • Transfer learning
  • OCR pipelines

Forecasting

  • Bi-LSTM
  • Q-learning (RL)
  • Time-series
  • Portfolio optimization
  • Statistical modelling

Data & Viz

  • Power BI
  • Tableau
  • Pandas / NumPy
  • Plotly / Seaborn
  • Excel (advanced)

Deployment

  • Flask / Streamlit
  • AWS SageMaker
  • DigitalOcean
  • Docker
  • Git / GitHub

Soft

  • Domain discovery
  • Stakeholder workshops
  • Technical writing
  • Research grounding
  • Async collaboration
How we work

A process biased toward shipping.

We don't run six-month discovery phases. Most engagements start with a paid 1-2 week scoping sprint that ends with a concrete go/no-go decision and a fixed-scope proposal.

01
Scoping sprint
1-2 weeks. We embed lightly, audit your data, understand the actual problem behind the brief, and return a written scope with timelines, costs, and the failure modes we see. Worst case: you walk away with a clear technical assessment.
02
Prototype
2-4 weeks. A working prototype against your real data — not a synthetic demo. The goal is to prove the ML approach works on *your* problem before committing to production engineering.
03
Production
4-12 weeks. Hardened code, deployment, monitoring, documentation, and handover. We write code your engineering team can maintain — no black boxes.
04
Handover & support
Optional retainer for the first 3-6 months post-launch. Model drift, edge-case handling, and lightweight training for your in-house team if you have one.
Get in touch

Have a problem worth solving?

If you're sitting on a dataset that isn't being used, an ML pilot that stalled, or a process that could be automated with vision or language models — let's talk. 30-min calls are free.

info@kratvanta.com →
Or book a slot directly · cal.com/kratvanta