Efficient Domain Adaptation via Instruction-Tuned LLM Fine-tuning at Scale
M. Babbar, K. Madhok, P. Barrach-Yousefi
NeurIPS Workshop on Foundation Models · Mar 2025 · 14 pages
AI Research Lab & Product Studio
verJSON is a research-driven AI lab that ships real products. From LLMs to autonomous pipelines.
AI isn't a feature.
It's the foundation.
We don't bolt on intelligence.
We architect it from day one.
Every product begins in the lab. We don't ship until the science is solid.
We build for real industries like agriculture, logistics, fashion, and finance, not generic use cases.
We co-found products, not just code them. Skin in the game, always.
M. Babbar, K. Madhok, P. Barrach-Yousefi
NeurIPS Workshop on Foundation Models · Mar 2025 · 14 pages
P. Barrach-Yousefi, K. Madhok, M. Babbar
arXiv preprint · Jan 2025 · 28 pages
M. Babbar, K. Madhok
ACL Findings · Nov 2024 · 11 pages
K. Madhok, P. Barrach-Yousefi
EMNLP Main Conference · Sep 2024 · 16 pages
M. Babbar, K. Madhok, P. Barrach-Yousefi
ICML Workshop on Spurious Correlations · Jul 2024 · 9 pages
P. Barrach-Yousefi, K. Madhok, M. Babbar
arXiv preprint · May 2024 · 19 pages
Campaigns, automated replies, and real-time performance analytics for WhatsApp at scale.
Automate workflows, decode textile specs, and coordinate vendors in one smart workspace.
AI-driven property research, analysis, and transaction automation for modern real estate.
Prompt-to-page navigation with on-page guidance, markup, and guided chronology inside any site.
End-to-end shipment visibility with predictive ETA and anomaly detection across carriers.
AI-driven dashboards and dynamic knowledge transfers that kill project miscommunication.
A single-window platform for farmers, NGOs, and researchers to access tailored solutions and markets.
Collaborative schema management. Propose, review, and version database changes with agentic AI for duplicate detection, risk analysis, and best practices.
A connected platform for human rights organizations to coordinate advocacy, share signals, and track cases across jurisdictions.
We embed with domain experts to uncover pain points that generic AI solutions miss. Two weeks of field research before a single line of code.
Model selection, data pipeline planning, and system design happen in parallel. We choose the right tool, not the trendiest one.
A working MVP in 30 days. We test with real users in the target domain before hardening the system.
Reliability engineering, monitoring, and edge-case handling. We build for the 1% scenario as carefully as the 99%.
Post-launch data loops back into model improvement. Every user interaction makes the product smarter.
We stay until the product can stand alone, with trained operators, runbooks, and a clear growth roadmap.
MLOps & Workflow Automation
Scale and optimise your machine learning processes efficiently.
Business Insights & Classification Models
Gain actionable insights through advanced AI models.
LLM Fine-tuning
Tailor large language models to fit your unique business needs.
Vision & Browser Automation
Streamline tasks with automated visual and web interactions.
Exploratory Data Analysis (EDA)
Simplify data integration with automated extraction, transformation, and loading.
ETL Pipelines
Automate data extraction, transformation, and loading for seamless integration.
From soil sensors to supply chains. AI that understands what grows and why.
Spec decoding, vendor coordination, and trend intelligence for fabric traders.
Property intelligence, market signals, and transaction automation at scale.
End-to-end shipment visibility with predictive ETA and anomaly detection.
Knowledge management, communication intelligence, schema governance, and workflow orchestration.
Coordinated advocacy, signal sharing, and case tracking across jurisdictions.
Designed a visual, modular AI agent system using OpenAI's Agent SDK, where agents can be composed using drag-and-drop blocks for tasks such as decision-making, research, and action execution. Integrated with PydanticAI and OpenAI/Groq LLMs, supporting parallel tasking and workflow memory.
60%
Reduction in manual task setup for clients
85%
Reusability of agent blocks across workflows
Built an AI-based monitoring system where agents track metrics (e.g., revenue, usage, error spikes, downtime) across LLM applications and raise alerts when anomalies are detected. Supports multi-agent setups with Prometheus and LLM-based log summarizers.
40%
Fewer missed alerts via smart triggers
80%
Faster incident triaging with summaries
Created a voice agent that listens for commands and controls industrial IoT devices. Integrated OpenAI Whisper for transcription, command parsing with LLMs, and MQTT communication with edge devices for real-time action execution.
98%
Command recognition accuracy in live tests
50%
Faster response vs. manual control
Messages automated monthly
via WAAutomate across 40+ brands
Farmers onboarded
through Kisaan Companion partners
Industries served
agri, logistics, fashion, finance, real estate, SaaS
Carrier APIs integrated
with unified webhook events via Postzit
Hallucination reduction
on domain-specific RAG pipelines
From brief to MVP
our standard delivery cycle

Pouya Barrach-Yousefi
Co-Founder
AI/ML leader bridging technical innovation and business strategy across pharma, healthcare, and SaaS.

Manish Babbar
Co-Founder
AI strategist combining machine learning, economics, and large-scale analytics into shippable products.

Kartikaye Madhok
Co-Founder
Data scientist focused on turning complex business problems into clean, deployable models.

Aditya Sharma
Product Designer
Human-centered designer specializing in B2B and enterprise interfaces that scale with users.

Karan Panchal
Full-Stack Developer
Builds responsive applications, owning frontend, backend, and the glue between them.

Ajit Gadkari
Full-Stack Developer
Ships interactive, high-performance web applications end-to-end with a bias for speed.
_get_in_touch
verJSON is a research-driven AI lab that ships real products. From LLMs to autonomous pipelines.
Response time
< 24 hours
Engagement types
Build · Consult · Co-found
Based in
India · Remote-first