ML engineer working the full path: raw data → modeling → evaluation → systems.
I'm Reyyi Shreyas, an AI/ML engineer studying at BMSIT&M, president of ASTRA (BMSIT&M's defence-technology student organization) and an ML intern at Launched Global. I like problems where a model has to survive real data, not just a notebook.
Most of what I learn happens in public: recent code, CI and documentation work for nilearn, PCNtoolkit, movement and mcp-mifosx, plus applied projects in deep learning, ensemble methods and agentic LLM systems.
This page tells that story in order: the pull requests other teams merged, the systems I've built, the research I'm pushing forward, the roles I've held, and the way I work.
I'd rather read the source than trust the demo.
Currently open to AI and ML internships.
- Researching lightweight hallucination detection for quantized small LMs
- Building and evaluating agentic LLM systems
- Contributing to open-source scientific Python and MCP server projects
- Exploring efficient ML systems and evaluation
First, proof in other people's repositories: pull requests maintainers reviewed and merged.
| Repository | Contribution | Status |
|---|---|---|
| mcp-mifosx #530 | Fixed stale paths and commands in README and CONTRIBUTING |  |
| mcp-mifosx #529 | Added pytest execution to Python CI, which installed it but never ran the tests | |
| mcp-mifosx #528 | Applied domain-aware action lookup in the suggestion engine, fixing empty results for the loans domain | |
| mcp-mifosx #522 | Matched plural keywords in intent routing so status queries return the right actions | |
| nilearn #6614 | Fixed the decoder training message being logged multiple times per fit |
|
| nilearn #6597 | Fixed stale standardize docs across the user guide and docstrings; added a gallery example |
|
| PCNtoolkit #551 | Documented basis_column, nknots and degree in the HBR normative-modelling tutorial |
|
| movement #1115 | Added the IO-guide update step to the Implementing new loaders & writers guide | |
Open pull requests under review: PCNtoolkit #563 and #566, movement #1125, #1127 and #1140, mcp-mifosx #550, artem-is #164 and #165, mcp-mifosx-self-service #43 and #44, sentiment module #14.
Next, what I build: five projects, each carried from a first model to a working system.
|
ChessMind AI Adaptive chess trainer using a two-model shared-memory architecture: a suggester LLM reasons over the position while a feedback LLM coaches from verified engine facts. Reduced first-feedback latency from 7.6s → 1.7s while keeping generation grounded in verified engine analysis. Includes a per-move Elo ensemble and reproducible evaluation harness.
|
Research Paper Analyst Agentic RAG that autonomously chooses between retrieval, summarization and comparison tools over a paper corpus, with grounded, source-cited answers, plus a RAGAS evaluation harness over a fixed question set.
|
|
Lunar image registration Sub-pixel registration of Chandrayaan-2 imagery to LRO NAC using SIFT correspondence and RANSAC, with a 7-module super-resolution pipeline (isro_1m) for hazard mapping.
|
Astra Chronos AI A 2-layer GRU predicts UAV trajectories from flight history; predictions drive threat scoring and autonomous interceptor guidance in a real-time 3D simulator.
|
|
TRICP Stacking-ensemble churn prediction that surfaces the factors behind each prediction, then triggers personalized retention workflows automatically.
|
Then, where I'm going deeper.
Researching a lightweight hallucination detection pipeline for quantized small language models using retrieval grounding, embedding-overlap signals and generation confidence.
The system performs sentence-level detection using a lightweight classifier and evaluates the trade-offs between FP16, 8-bit and 4-bit quantization across detection quality, hallucination rate, latency and GPU memory.
LLMs RAG Quantization Hallucination Detection Evaluation
Status: Preparing submissions to research conferences
Beyond the code: the roles behind the commits.
ASTRA: President BMSIT&M's defence-technology student organization. Leading technical, research, recruitment and cross-team initiatives.
ML Intern: Launched Global Worked on machine-learning applications and applied AI workflows.
Open-source contributor Contributed documentation, bug fixes and tooling improvements across scientific Python and open-source ML projects.
For scanners: the tools behind the work above.
| LANGUAGES | Python |
| ML / DEEP LEARNING | PyTorch | scikit-learn | XGBoost | LightGBM | OpenCV |
| LLM / RETRIEVAL | Transformers | LangGraph | LangChain | FAISS | RAGAS | MCP |
| ENGINEERING | FastAPI | Flask | Streamlit | Docker | pytest | Git |
| DATA | NumPy | pandas |
Understand → Experiment → Break → Ship. Read the implementation, understand the assumptions, build alternatives, measure what changes, test where systems fail, then turn what survives into something usable.
A model that works in a notebook is an experiment.
A model that survives evaluation, integration and real usage is a system.
The profile above, rebuilt as one scroll-lit journey: scenes, ambient sound and the work in motion.
AI/ML | Research | Open Source