My path threads together computer vision deployment, recommendation systems, and backend API ownership, consistently landing on the automation layer that removes manual or repetitive work for others.
ms latency
Specialized Functions & Roles.
Engineering capabilities and specialized roles demonstrated across real production work.
Machine Learning Engineer
AI Engineer
MLOps Engineer
Backend Engineer
Live in Production.
Each signal ties back to a specific, evidenced moment.
Built
Systems shipped end-to-end, from schema to deploy.
- •Published IEEE paper on CNN-based analysis of PCG signals
- •Built Automation Fixer to cut manual QA root-cause work by ~80%
- •Built a recommendation system with OpenAI embeddings and XGBoost
Learned
New tools and depth picked up on the job.
- •Published AI pathology research with Taylor & Francis
- •Experimenting with RAG and context grounding
Solved
Gaps closed - false positives, edge cases, dead ends.
- •Fixed a priority-client production OOM in the journal search service
Scaled
Taken from prototype to real production volume.
- •Deployed a YOLOv11 auto-cropping model to TruthIn's 1M+ user base
Juhi's career, mapped.
A timeline of milestones captured directly from real work.
Started applying AI to scientific imaging
Took ownership of ML pipelines in production
Built a recommendation system end to end
Moved into backend systems at scale
Resolved a production memory incident
Started building internal AI tooling
Daily tools.
What Juhi actually reaches for, pulled from the tags attached to captured moments.
Key Work By Role
What they built, shipped and owned at each role, with impact numbers where there are any.
AlgoSec
Fixed a priority-client production OOM in the journal search service
Diagnosed and fixed an out-of-memory issue for a priority client in /journal/v1/overview/search by introducing batched retrieval in Metro Facade, bounding per-request memory for large result sets.
Built Automation Fixer to cut manual QA root-cause work by ~80%
Built Automation Fixer, a tool that automates QA root-cause analysis and PR generation, and an Ollama/n8n RAG assistant over internal code and docs to support the Support team.
More from this role
- Developed REST APIs in ACE Network Analyzer Core (MS-Metro, Metro Facade API layer, Device Manager), the backend computing Journal change history and risks for ACE Network Analyzer; Kafka pub/sub between the two.
- Migrated NAC services from Java 11 to 17, fixing library and inter-service compatibility issues.
- Sanity-tested 7 releases (Dev, Stage, Production) via CI/CD on Kubernetes; reviewed 45% of team PRs.
Natfirst Pvt. Ltd.
Deployed a YOLOv11 auto-cropping model to TruthIn's 1M+ user base
Deployed a YOLOv11 auto-cropping model achieving 98% mAP@50 at 2.51 ms latency for TruthIn, an Android app with 1M+ users.
Built a recommendation system with OpenAI embeddings and XGBoost
Built a recommendation system combining OpenAI embeddings and XGBoost, reaching 95.73% Top-1 and 99.02% Top-5 accuracy, alongside an agentic evaluation workflow using LangChain, LangGraph, and CrewAI.
Independently built and deployed StudyBuddy, a full-stack AI tutor grounded in uploaded PDFs
Independently built and deployed StudyBuddy, a full-stack AI tutor that answers only from uploaded PDFs with page-level citations and adaptive exams, using an Express 5 REST API with an OpenAPI contract generating Zod validation and typed hooks.
More from this role
- Built a scalable ETL pipeline on AWS SageMaker, EC2 and S3 for automated training and model versioning (Docker).
IRDE, DRDO
- Built a Flask app automating aberration correction in hyperspectral images, cutting manual processing by 95%.
- Led data annotation and preprocessing (Label Studio) for an AI landmine-detection initiative with IIT Hyderabad.
Side projects, open source & publications.
Independent builds, open-source contributions, technical writing, and experiments beyond the day job.
StudyBuddy
Independently built and deployed a full-stack AI tutor answering only from uploaded PDFs, with page-level citations and adaptive exams; Express 5 REST API with an OpenAPI contract generating Zod validation and typed hooks. Benchmarked 5 vision models (Gemini Flash, Qwen2.5-VL 3B/7B, Gemma 3 12B) for PDF page reading: hosted Gemini was 7-30x faster per page (6.9 s vs 46-209 s) than local open-weight models; built a 72-item eval set.
RAG Experimentation
Experimenting with retrieval-augmented generation (RAG) techniques, including context grounding.
Sustainable AI Research
Researching sustainable AI as an area of ongoing study.
Deep Learning for Cardiovascular Health: CNN-Based Automated Analysis of PCG Signals
Authored a peer-reviewed paper on using CNN-based deep learning to automatically analyze phonocardiogram signals for cardiovascular health assessment, published with IEEE.
Formal record.
- M.Tech. (CS&E)2025 - PresentIndian Institute of Technology, Patna
- B.Tech. (CS&E, Hons: AI & ML)2021 - 2025Amity University UP
- Cisco® CCNAv7Cisco
- Scrum Fundamentals (SFC™)
- SQLHackerRank
- GitHub FoundationsGitHub
- Claude 101
Profile last updated 6 days ago.

