Applied AI research
Conformers, Vision Transformers, and Liquid Neural Networks. Implemented, benchmarked, and evaluated.
AI ENGINEER & PRODUCT BUILDER
Taranpreet
Singh Saggu.
From an idea to an
intelligent, working product.
I build AI systems that connect research with real-world use. Agentic workflows, deep learning, and the engineering that brings it all together.
200+
Users in month oneGurmat Darbar
80+
Samagams worldwideCommunity impact
IIT Mandi
AI research internshipMultimodal deepfake detection
Research ↗
Through to deploymentModels, APIs & cloud
01 / A LITTLE ABOUT ME
Curiosity in research.
Ownership in execution.
I’m an AI engineer and co-founder of Gurmat Darbar. I enjoy understanding how a model works, testing where it falls short, and building the complete system around it.
My work spans multimodal deepfake detection at IIT Mandi, agentic analytics, financial document retrieval, and local code generation. Alongside building, I’m pursuing degrees in data science and computer science.
Intelligent systems
MCP tool orchestration, agentic RAG, and local LLM pipelines that turn information into useful workflows.
End-to-end delivery
From frontend interfaces and secure APIs to cloud deployment and a community of real users.
02 / SELECTED WORK
Ideas, built into systems.
Four projects across agentic AI, retrieval, local LLMs, and neural network research.
AGENTIC AI / DEVOPS01
MCPilot
Agentic Analytics Platform
Ask questions across repositories, builds, deployments, and incidents. An AI-powered analytics platform that brings DevOps context into one conversation.
Natural-language queries across 5 repositories and 145 operational records.
Implementation details
- Built MCP-based tool orchestration across 5 repositories, 60 builds, 35 deployments, 20 incidents, and 30 issues.
- Used Gemini with Groq/OpenAI fallback for resilient LLM execution, and integrated the GitHub API through MCP for live repository, issue, and pull-request analysis.
- Built a Vue 3 + TypeScript frontend with multi-chat conversations, contextual execution traces, and operations dashboards.
AGENTIC RAG / DOCUMENT INTELLIGENCE02
StockLens
Financial Policy Intelligence
An agentic retrieval system for navigating Ministry of Finance and RBI policy documents, using semantic search to make financial publications easier to explore.
Semantic retrieval over 2024 monthly financial reports.
Implementation details
- Built an agentic Retrieval-Augmented Generation system to analyze Ministry of Finance (MoF) and RBI policy documents.
- Indexed 2024 monthly financial reports with vector embeddings for semantic search over government publications.
- Implemented agentic workflows with Redis caching and vector search to improve query efficiency and response latency.
LOCAL LLMS / DEVELOPER TOOLS03
Autocoder
Local Code Generation System
A local, fine-tuned LLM pipeline that brings code generation directly into Jupyter. Prompt, stream, and insert generated Python into notebook cells.
From a %%prompt command to executable notebook code.
Implementation details
- Configured local .gguf models in Ollama using a custom ModelFile for a fine-tuned LLM pipeline.
- Developed a Jupyter magic command (%%prompt) to stream responses and automatically insert generated Python into notebook cells.
- Implemented an end-to-end workflow with CLI testing and notebook integration for interactive code generation and execution.
DEEP LEARNING / TIME SERIES04
Hand Gesture Segmentation
Liquid Neural Networks Research
Reproduced Liquid Time-Constant networks from scratch to segment hand gestures, and benchmarked their temporal adaptability against an LSTM baseline.
54.6% validation accuracy with a 23% lower generalization gap.
Implementation details
- Explored Liquid Neural Networks, a continuous-time recurrent architecture using Liquid Time Constants for temporal adaptability.
- Reproduced the LTC network (Hasani et al., 2020) and implemented hand gesture segmentation on 32-dimensional time-series motion data.
- Built custom LTC cells in PyTorch: 54.6% validation accuracy versus 45.0% for LSTM, with a 23% lower generalization gap.
03 / EXPERIENCE
Learning deeply. Building deliberately.
Co-Founder
Gurmat DarbarCo-founded and developed a full-stack platform connecting the community through Samagams, events, speakers, and registrations.
- Reached 200+ users in the first month and supported 80+ Samagams worldwide.
- Designed frontend interfaces and scalable backend APIs with secure authentication and role-based access.
- Led deployment and infrastructure management on Google Cloud Platform for scalability and reliability.
AI Research Intern
IIT Mandi
Researched multimodal deepfake detection, bringing audio and visual signals together in end-to-end experimental pipelines.
- Explored Conformer architectures for audio and image spoof detection.
- Integrated Vision Transformers (ViT) to classify bonafide and spoof inputs with improved generalization.
- Built PyTorch pipelines and optimized preprocessing, feature fusion, and model evaluation.
04 / THE TOOLKIT
The tools behind the work.
A foundation in computer science, with a focus on practical AI engineering.
Languages & fundamentals
Machine learning & AI
Frameworks & tools
Also used in my projects
MCP / PostgreSQL / Vue 3 / TypeScript / Ollama / GCP
05 / EDUCATION
Two paths. One foundation.
2022 — 20277.0 CGPA
Indian Institute of Technology, Madras
BS in Data Science and Applications
Data science & applications2023 — 20279.0 CGPA
Guru Tegh Bahadur Institute of Technology
B.Tech in Computer Science (AI & ML)
Guru Gobind Singh Indraprastha University (GGSIPU)LET’S CONNECT
Have an interesting
problem in mind?
For collaborations, opportunities, or a conversation about AI.
I’d love to hear what you’re working on.