Shashwat Saket
ML / AI Engineer
ML/AI engineer with 2+ years at CGI Inc., where I led a cross-functional team of 8 shipping conversational-AI modules, ML anomaly detection on financial data, and Kafka + Spring Boot microservices (Credit Studio Star Award, Q1 2025). Now MS CS @ NYU (3.56 GPA) and Graduate Researcher in the Varol Lab — building hierarchical codebooks for two-tier ANN retrieval, LLM knowledge-graph RAG, and neural decoding models that took 1st/24 in a Kaggle benchmark.
Research Playground
Play with the actual mechanics behind my projects — quantization tradeoffs, connectome structure, decoder drift, scaling curves, and grounded retrieval.
Index cost: 50–1300× cheaper than training two codebooks.
Start from the finest level and drop centroids. Penalty stays near zero at the fine tier — the practical winner.
Projects
Active research systems and engineering artifacts.
ConnectionMiner — Connectome × Transcriptomics
Neuroinformatics platform (NYU Varol Lab) integrating single-cell RNA-seq transcriptomics with the FlyWire connectome to map neuronal identity across 741 neuron types in the Drosophila visual system.
Reusable Codebooks for Two-Tier ANN Retrieval
Hierarchical quantization framework enabling cheap ANN prefiltering and expensive reranking from a single shared codebook. Matches Product Quantization accuracy while reducing index construction cost by 50–1300×.
Neural Decoding — Intracortical Kaggle Competition
22-model ensemble predicting finger kinematics from 96-channel intracortical recordings across 312 sessions, reaching 0.7373 R² under multi-year neural drift.
ReservaDirect — Autonomous Reservation Agent
AI concierge that turns Google Calendar into a reservation system: it phones restaurants directly, confirms the booking, and writes the result back to your calendar.
Spatio-Temporal Crop Prediction with Metaheuristics
Crop prediction system using metaheuristic feature optimization — cuckoo search, PSO, and ant colony optimization — over spatially and temporally varying agricultural data.
Conversational AI Modules – Credit Studio
Three NLP-based conversational AI modules automating 40% of repetitive tasks and increasing user engagement by 10%.
ML Anomaly Detection on Financial Data
Elasticsearch ML-based anomaly detection on large financial datasets, improving detection accuracy by 25%.
Experiment Timeline
Tracking architectural evolution and performance gains.
Erica — Initial RAG Pipeline
Basic vector embedding retrieval with local LLM inference. Functional but prone to hallucination on multi-hop questions.
Neo4j Knowledge Graph Integration
Added Neo4j knowledge graph for structured reasoning paths. Hallucination reduced significantly on domain queries.
Semantic Extraction Pipeline
LLM-driven extraction of concepts, definitions, and prerequisite relationships from unstructured web/multimedia data.
End-to-End ML Pipeline
Full ML pipeline for ingestion, chunking, embedding, inference, and async execution. Supports long-form instructional outputs.
Across 4 iterations: knowledge graph grew 1,000 → 5,000+ nodes, retrieval moved from vector-only to hybrid RAG, and the stack reached a production FastAPI + Docker pipeline
Publications
Peer-reviewed contributions to the field.
Security provisions in smart edge computing devices using blockchain and machine learning algorithms: a novel approach
Shashwat Saket, et al.
Proposed a novel security framework combining blockchain and ML for smart edge computing. Published in Springer Cluster Computing (Q1, IF 4.1).
Travelling Guidance Using ACO and HBMO Techniques in COVID-19 Pandemics: A Novel Approach
Shashwat Saket, et al.
First-author paper applying Ant Colony Optimization and Honey Bee Mating Optimization to COVID-19 travel route planning, reducing route costs by 15%.
Hybrid Approach for Deception Tracing in Smart Cities Using LR and n-fold Intelligent Machine Learning Techniques
Shashwat Saket, et al.
Developed a hybrid ML approach combining Logistic Regression with n-fold techniques for deception tracing in smart city environments.
AI System Design
Architecture blueprints for production ML systems.
Production retrieval-augmented generation with hybrid search, reranking, and streaming output.
Research Direction
What I'm Exploring
- Connectome × transcriptomics alignment across 741 neuron types in the Drosophila visual system
- Nested codebooks that serve both tiers of an ANN retrieval pipeline
- Neural decoding that stays accurate under multi-year electrode drift
- LLM-driven knowledge graph reasoning with RAG and Neo4j
Open Problems I Care About
- How much of a neuron's molecular identity is recoverable from connectivity alone?
- Can quantization schemes trade MSE for recall in a way we can actually prove?
- How do decoders adapt to distribution shift without labelled target data?
- Can RAG systems provide formal retrieval guarantees instead of empirical ones?
Research Vision
- Treating brains and retrieval systems as the same problem: structure that has to be recovered from partial, noisy measurements
- Building ML systems that are production-ready, interpretable, and robust to domain shift
- Keeping theory and benchmarks in the same loop — theorems that predict what the experiment shows
Research Journey
From theoretical curiosity to building intelligent systems.
The Foundation — B.Tech at BIT Mesra
Started Computer Science with a specialization in Computational Intelligence. Explored data mining, cloud security, and published early research in Springer journals.
ML Internship & First Publications
Applied meta-heuristic optimization to real-world datasets. Published first-author paper on travelling guidance using ACO and HBMO techniques at Springer LNNS.
Industry — Software Engineer at CGI Inc.
Led ML + backend teams building Credit Studio. Shipped conversational AI modules, event-driven microservices with Kafka, and ML-based anomaly detection on financial data.
NYU — MS CS & Graduate Researcher, Varol Lab
Graduate studies at NYU alongside computational neuroscience research: ConnectionMiner maps neuronal identity across 741 Drosophila neuron types by fusing single-cell RNA-seq with the FlyWire connectome. Also building Erica (LLM knowledge-graph RAG) and hierarchical codebooks for two-tier ANN retrieval.
HOF — a classic NYU hangout
Washington Square Arch — the heart of NYU
Trophy in hand at the NYU gym
Snowy morning walk to Courant
Broadway blocks around the NYU campus
Manhattan skyline from Brooklyn Heights
Times Square lights
Midnight crowds in Times Square
Radio City Music Hall after dark
Research Trajectory
Graduate Researcher, Computational Neuroscience
New York University — Varol Lab
Architecting ConnectionMiner: integrating single-cell RNA-seq transcriptomics with the FlyWire connectome across 741 neuron types in the Drosophila visual system, via an iterative matrix-factorization solver reaching r = 0.54 at the 75-cluster scope.
MS Computer Science (in progress)
New York University — Tandon School of Engineering
GPA 3.56/4.0. Coursework: Design and Analysis of Algorithms, Artificial Intelligence, Machine Learning, Neuroinformatics — where six invitation-only Kaggle tracks produced a 1st/24 finish on intracortical neural decoding.
Software Engineer
CGI Inc.
Led a cross-functional team of 8 ML + backend devs. Shipped 3 conversational AI modules using NLP, built 10+ Camunda-based workflows with Spring Boot & Kafka, and configured ML anomaly detection on financial datasets.
Machine Learning Intern
Birla Institute of Technology
Applied meta-heuristic optimization (ACO, HBMO) to COVID-19 datasets. Built NLP-based propaganda detection model with 25% improved detection rate.
B.Tech Computer Science & Engineering
Birla Institute of Technology, Mesra
CGPA: 8.73/10. Specialization in Computational Intelligence (9.30/10). Published 4 papers in Springer & Enderscience journals.
Get in Touch
Open to research collaborations, ML engineering opportunities, and internships.
© 2026 · Built with precision · All systems nominal