Preetham Noel P
AI Product Engineer
Agentic AI · LLMs · Data Science
2× Patent Holder · 3× Published Researcher · AI Systems · 4 Startups
About
Building AI that actually ships — currently deep into agentic systems, LLM evaluation, and full-stack AI product engineering. I've taken AI projects from early prototypes to live products across five companies, and I enjoy both the research depth and the engineering discipline it takes to make them work in the real world.
My path into AI
My interest started during undergrad at Christ University (Data Science + AIML Honours, 9.82 CGPA), where curiosity about intelligent systems quickly became hands-on work. I began building, publishing, and filing patents before finishing my degree — and haven't stopped since.
Research to production
I hold 2 patents filed with Intellectual Property India — one in healthcare optimization, one in AI-driven cyberbullying detection using a Hybrid LSTM-BERT architecture. Alongside that, 4 international research papers published through Springer and John Wiley & Sons. Research gave me rigour; product work gave me speed.
What I've shipped
Across five companies, I've built full-stack AI systems — from database design and API architecture to LLM evaluation pipelines and automation tooling that measurably reduced manual workflows. Outside commercial work, I contributed through my university's service learning programme by building a web platform with ML-backed analytics for an orphanage, covering donation forecasting, volunteer engagement classification, and real-time activity tracking.
Building the community
Served as Head of Data Science at ASCII, Christ University — organised 10+ industry workshops, trained 100+ students, and mentored 30+ members in applied ML and data science.
What I'm exploring now
Multi-agent orchestration, structured LLM evaluation frameworks, agentic workflow automation, and the engineering decisions that make AI products reliable at scale.
Let's connect
Open to AI product and data science roles. Always happy to discuss interesting problems, potential collaborations, or what it actually takes to ship AI.
Work Experience
- Designed and built the database schema and API layer for Lipi (LLM-powered travel agent), enabling seamless communication between AI backend and frontend; integrated and tested TBO travel APIs.
- Implemented multi-agent orchestration with intent-aware routing, enabling dynamic handoffs between specialized agents (flights, hotels, itinerary planning) while preserving context across user workflows.
- Built a PR automation pipeline using Python and GitHub Actions, and led architecture design for an SDLC AI Agent system, defining backend, APIs, testing, and prompt workflows for lifecycle automation.
- Mapped user personas, designed UI flows, and defined interaction scenarios for Lipi across 15+ user journeys, used directly as the basis for engineering requirements.
- Produced demo videos for the initial product walkthrough and helped bridge product concepts into technical specs, enabling smoother handoff between design and development.
- Integrated OpenAI Evals for automated end-of-day LLM benchmarking, improving overall interview experience quality by 30%.
- Built an MCP-driven retrieval tool automating manual data workflows for the account management team, reducing manual effort and processing time by 50%.
- Built and designed APIs for a FastAPI-based interview anomaly detection and reporting platform, extending existing APIs with additional parameters and developing new endpoints for flagging session irregularities.
- Built an AI agent for automated interview question generation, reducing manual question curation effort for the team.
- Prototyped a Hugging Face TTS pipeline to explore feasibility of an AI-driven music and lyrics generation feature.
- Designed a MySQL + MongoDB hybrid storage layer for the song selection model, optimising for mixed structured and unstructured data access patterns.
- Researched information extraction and retrieval approaches to support data integration across the sustainability ecosystem, contributing findings to internal research workflows.
- Assisted in building internal data tooling to streamline sustainability reporting, supporting the team in aggregating and analysing clean tech data.
- Built an automated issue ticketing system using GPT and token chunking to classify issue types and route them to the appropriate support channels, reducing manual triaging effort.
- Fine-tuned GPT-4 prompt chains using variance tracking, improving LLM output consistency by ~18% and reducing manual review of AI-generated content.
Research & Publications
Research papers, patents, and academic contributionsOptimizing Healthcare Towards a Beneficial Future with Cutting-Edge Medical Solutions
Patent · 2024Intellectual Property India
Patented healthcare optimization system leveraging cutting-edge medical technologies to improve patient outcomes and operational efficiency. As part of the patent application process, a structured patient feedback survey was conducted at hospital registration desks to validate real-world applicability and ground the solution in observed clinical needs.
AI-Driven Cyberbullying Detection and Intervention System Utilizing Hybrid LSTM-BERT Architecture
Patent · 2024Intellectual Property India
A hybrid deep learning architecture combining LSTM and BERT for real-time cyberbullying detection and automated intervention across online platforms.
Performance Analysis of CPU and GPU Processors Using Advanced Data Analysis Techniques
Springer 2024MIT ETMCIS 2024 — International Conference on Emerging Trends in Microelectronics, Communication & Intelligent Systems
Comparative performance study of CPU and GPU architectures using advanced statistical and data analysis methods, presented at an international conference hosted at MIT.
Statistical Data Analysis of Anticorrosion and Antifouling: Unveiling Insights from Performance and Trends
Journal · Oct 2024John Wiley and Sons, Inc.
In-depth statistical analysis of anticorrosion and antifouling performance data, uncovering trends and actionable insights from large-scale experimental datasets.
Portable and Automated Healthcare Platform Integrated with IoT Technology
Journal · Aug 2024John Wiley and Sons, Inc.
Design and implementation of a portable, integrated platform enabling automated monitoring, data collection, and real-time health analytics.
Achievements & Recognition
2 Patents Published — Intellectual Property India
Both patents published in 2024, covering AI-driven cyberbullying detection (Hybrid LSTM-BERT) and a cutting-edge healthcare optimization system.
J.P. Morgan Hackathon — Shortlisted
Selected among top performers at the J.P. Morgan Hackathon — one of 5+ shortlists across 10+ national hackathons including Smart India Hackathon.
Head of Data Science — ASCII, Christ University
Led 10+ workshops and industry sessions, training 100+ students and mentoring 30+ members. Improved technical proficiency by 18% based on pre/post assessments.
NPTEL Project Management — Top 5%, IIT Roorkee
Achieved 83% and ranked in the top 5% of participants nationally in the Project Management certification by NPTEL, IIT Roorkee.
Meritorious Scholarship — Christ (Deemed to be University)
Recipient of the Meritorious Scholarship for academic excellence. Completed B.Tech in Computer Science (Data Science) with Honours in AIML, with a CGPA of 9.82 / 10.
Skills
Projects
AI-Powered Resume Screening & Interview Platform
Extended an existing hiring platform with candidate ranking and AI-powered interview question generation. Integrated OpenAI Evals to benchmark LLM output quality across screening tasks, enabling continuous performance monitoring and regression detection. Automated shortlisting logic reduced manual review to ranked candidate pools rather than raw applicant lists.
LANA — Load, Analyze & Advance
Local AI analysis tool built with FastAPI and Ollama — upload a CSV, Excel, or JSON file and query it in plain English, with the LLM running fully on-device. Automated cleaning pipeline handles duplicates, missing values, and outliers with user-confirmed fixes; analytics backend covers statistical profiling, linear regression, and chart generation with one-click export to PDF, Word, and CSV.
↗Intelligent Banking Assistant — South Indian Bank
NLP-driven platform enabling bank staff to query customer feedback in natural language, backed by a text classification pipeline for intent detection. Built automated alert triggers for real-time negative sentiment and complaint spike detection, plus a query-driven credit card eligibility assessment system.
Website for Orphanage with ML-Backed Insights
Responsive web platform built as a social cause contribution through the university service learning programme. Covers donation forecasting, volunteer engagement classification, and real-time activity tracking — production-grade engineering applied to community impact.
Open Source Contributions
Pull requests and issues across AI, ML infrastructure, and developer toolingfix(session): restrict session dir and file permissions to owner-only
PR · Mergedalibaba / open-code-review
Fixed a security vulnerability where session directories and files were created with world-readable permissions. Changed directory access from 0755 → 0700 and files from 0644 → 0600, preventing local users on shared systems from accessing sensitive session data including source code diffs and model analysis. Added a regression test to verify correct permissions on Linux/macOS.
docs(pages): add ocr scan documentation page
PR · Mergedalibaba / open-code-review
Added documentation for the ocr scan command to the project website — covering flag references, usage examples, preview mode, and batching behaviour. Resolves a tracked gap in docs for a core LLM-powered static code analysis feature.
[Anima] Add img2img pipeline blocks
PR · Openhuggingface / diffusers
Added image-to-image functionality to the Anima modular pipeline by introducing three new pipeline blocks integrated into AnimaAutoBlocks. Implementation follows the established pipeline pattern, allowing img2img mode to be triggered via an image parameter. Reviewer approved; 13 tests passing, 5 skipped.
Callable class middleware raises AttributeError instead of MiddlewareException
Issuemicrosoft / agent-framework
Identified a bug where callable class instances used as middleware trigger an unhelpful AttributeError because validation code accesses middleware.__name__, which class instances do not expose. Proposed fix: replace bare __name__ access with getattr(middleware, '__name__', type(middleware).__name__) at three locations in _middleware.py for clear error messaging.
Add XFORMERS_VERBOSE_DISPATCH env var to log selected attention kernel at runtime
Issuefacebookresearch / xformers
xformers silently selects its attention kernel (Flash3, Cutlass, Triton, or PyTorch fallback) with no runtime visibility. A driver update or CUDA upgrade can silently route calls to a 4–10× slower fallback with no log output. Proposed XFORMERS_VERBOSE_DISPATCH following the established patterns of TORCH_LOGS in PyTorch and JAX_LOG_COMPILES in JAX.
Education
Christ University, Bangalore
B.Tech in Computer Science (Data Science) · Honours in Artificial Intelligence & Machine Learning
2022 – 2026