Reports with a source behind every line
A RAG assistant drafts medical assessment reports, citing the document behind each statement so clinicians can check it.
−60%time to draft a medical report
See the role AI EngineerIndia
Systems you can check.
Building & Deploying Enterprise AI Systems. Tech Consultant specializing in agentic workflows, production RAG, and real-time vision systems.
60%
less time drafting medical reports<2s
to detect an accident on a live camera+25%
answer relevance over the internal baselineWhere it runs
Three places my work runs today, and what it changed there.
Live cameras that raise their own alarms
Accident, fire and congestion detection across a city's traffic cameras, from stream ingestion to the alert.
<2sfrom incident to alert
See the case study AI platforms moved, secured and made multilingual
An AI platform moved from Azure to Google Cloud, locked behind Entra ID and IAP, translating documents across Indian languages.
Azure → GCPon Cloud Run, end-to-end auth
See the role 01 · Work
01 · Research agent2026
Nodus
Trace every claim back to its paper.
Finds the papers, pulls out each claim, clusters them across papers, and scores every finding for lineage, disagreement and quality, with the score's workings shown.
02 · Multi-agent code review2026
ReviewBot
Nothing merges on silence.
A GitHub App whose merge check clears only when a named developer records a decision on every major finding. Not on a fix, on a decision. Live on real repos.
03 · Real-time vision2024
VideoAnalytics
Alerts come from sequences, not frames.
Accident, fire and congestion detection across a city's live traffic cameras. Alerts arrive in under 2 seconds without paging anyone over a single glitchy frame.
02 · Experience
Where I've shipped.
2024
2025
2026
AI EngineerCities
Associate Data ScientistHealthcare & logistics
Tech ConsultantEnterprise · Data & AI
Enterprise
Tech Consultant, Data & AI
Moving enterprise AI platforms to Google Cloud and making them secure and multilingual.
Azure → GCPplatform migrated to Cloud Run
I moved an AI commentary-generation platform from Azure to Google Cloud, containerising its FastAPI services and running them on Cloud Run with Cloud Build, Cloud SQL, Secret Manager and Artifact Registry. Every request is authenticated end to end: Entra ID handles tenant identity, and IAP/IAM decides which services and users get through. Alongside that, I built document translation for Indian languages with PyMuPDF and Azure OpenAI that keeps tables, figures and page layout intact.
Healthcare & logistics
Associate Data Scientist
GenAI assistants for healthcare and forecasting for logistics.
−60%time to draft a medical report
+25%relevance over the internal baseline
My main work was a RAG assistant that drafts medical assessment reports, with settings for each disease type and a citation behind every statement. It cut drafting time by 60%. A prompt-tuning pipeline that classifies each query and adjusts generation settings to match lifted relevance scores by 25% over the internal baseline. I also built air-cargo demand forecasting that combines economic indicators, shipment trends and live weather data.
Cities
AI Engineer
Real-time computer vision on live city traffic cameras.
−30%frame loss in stream ingestion
<2sto detect an accident
+18%routing efficiency in simulation
I built the computer-vision stack behind a city's traffic cameras. RTSP ingestion with GStreamer and RabbitMQ cut frame loss by 30%, and accidents and traffic anomalies are detected in under 2 seconds. A congestion heatmap for rerouting improved routing efficiency by 18% in simulation, and along the way I fine-tuned Transformer and YOLO models with Hugging Face and Ultralytics.
Read the VideoAnalytics case studyBuilt in India, running in production.
About
03 · About
A model's output is only useful if someone can see where it came from, and push back on it.
I'm an AI engineer with nearly three years of building GenAI and computer-vision systems that real people depend on: medical report assistants, city-wide traffic monitoring, and enterprise platforms on Google Cloud.
So I design for trust from the start. That means citations, scoring rules you can read, human sign-off, and an audit trail.
I care just as much about the unglamorous half: queues, retries, auth, deployment pipelines and cost. That's what makes a demo something a team can actually run.
Process
Understand it.
I start with your problem, not a tool. Who uses it, what it has to get right, and what failure would look like.
See what's there.
Before proposing anything, I look at the data you already have and the architecture underneath it, so the plan fits what's really there.
Research or propose.
If the problem is new to me, I research it first. If I've solved something like it before, I propose a direction straight away.
Map it out.
You get a roadmap or an architecture you can question: the moving parts, the risky bits, and the order to build them in.
Build it.
Then I build the final solution end to end, and hand it over running in production, not as a prototype.
04 · Toolkit
GenAI & LLMs 07
Agent graphs, retrieval, and prompts that cite their sources.
Vision & Deep Learning 05
Detection and tracking on live video.
Data Science 08
Analysis, forecasting and quick dashboards.
Backend & Data 13
The services, queues and stores underneath.
Cloud & MLOps 11
Shipping it: containers, identity and pipelines.