Trisynapse Memory
Agent Memory Engine
A local-first, high-performance memory architecture and engine designed for AI agents, multi-agent workflows, and local LLM research systems.
Chennai, Tamil Nadu, India
Shanmukha Sainath
AI Engineer · KLA Corporation
I'm Shanmukha Sainath, AI Engineer at KLA Corporation. I specialize in building algorithms for high-precision optical metrology, and I love coding models from scratch to understand them under the hood. Here, I document my experiments, research papers I read, and the systems I build across NLP, Computer Vision, RL and Generative AI.
Active Projects
Trisynapse
Desktop AI Assistant
A local-first, privacy-focused desktop AI research assistant. It integrates FastAPI, Electron, React, and LanceDB to let researchers build local knowledge graphs, run semantic search, and organize notes.
AresSim
Mars Environment Simulator
A Mars habitat development environment where I am testing different RL Algorithms and LLM Agents to perform the task of building a habitat using a rover as an agent.
LLM Gathering
Multi-Agent Platform
A platform where different LLMs with assigned roles discuss any topic that is given. Helps analyze reasoning behavior across LLMs and build Agent Memory engines for multi-agent workflows.
Experience
March 2026 – Present
Chennai, India
AI Engineer 3
- Researching and developing algorithms for next-generation optical metrology tools.
- Focus on integrating physics modelling and AI for semiconductor inspection.
July 2023 – March 2026
Chennai, India
AI Engineer 2
- Built deep learning models for optical metrology to achieve sub-nanometer (angstrom-level) measurement precision in semiconductor manufacturing.
- Designed and implemented synthetic data generation pipelines, boosting model robustness and outperforming legacy methods.
- Collaborated with engineering teams to deploy, profile, and optimize algorithms for high-throughput production tools.
May 2022 – July 2022
Chennai, India
Machine Learning Intern
- Improved semiconductor defect classification F1-score by 3% on heavily imbalanced datasets.
- Experimented with self-supervised learning (SimCLR), GANs, and feature-generation networks for data augmentation.
- Optimized training pipelines using class-weighted losses (Focal Loss) and active sampling.
January 2022 – April 2022
Remote
Applied Scientist Intern
- Integrated Vision Transformers (ViTs) into image-based ad moderation models, boosting accuracy and deploying to production.
- Researched few-shot adaptation of CLIP using prompt-tuning and cache-based contrastive learning.
- Improved defect detection F1-score by 6% through fine-tuning and model compression.
Earlier research and engineering roles · 2
Deep Learning Research Intern
Deakin University
- Researched split-learning architectures to train neural networks across multiple client devices while maintaining data privacy.
- Implemented Manifold Mix-up for secure activation transfer between client and server.
- Analyzed client capacity limits, Dirichlet distributions for non-IID data, and gradient noise trade-offs.
Machine Learning Intern
ResoluteAI.in
- Built and deployed a real-time face recognition model for biometric attendance systems.
- Implemented custom triplet and contrastive loss functions in TensorFlow to train Siamese networks.
- Used t-SNE and K-Means to analyze and visualize face embeddings.
Education
B.Tech, Electronics & Electrical Communication Engineering
IIT Kharagpur
Minor in Computer Science Engineering
Micro in Artificial Intelligence and Applications
Projects
Showing 8 entries
project
May 2026 – Present
Trisynapse Memory
AI Agent Memory Architecture & Engine
Selected workIn progress
A local-first, high-performance memory architecture and engine designed for AI agents, multi-agent workflows, and local LLM research systems.
PythonTypeScriptVector DBLLM AgentsMemory
project
April 2026 – Present
Trisynapse
Desktop AI Research Workspace
Selected workIn progress
A local-first, privacy-focused desktop AI research assistant. Bring your own keys to compile sources into local wikis, explore concept graphs, perform semantic search, utilize an infinite whiteboard canvas, and monitor topics offline.
ElectronReactFastAPISQLiteLanceDBExcalidraw
project
February 2026 – April 2026
DotChat (SLM Framework)
Causal Language Model & ChatUI
Selected work
A minimalist PyTorch framework and web interface for training small language models (DotLM) from scratch. Supports pretraining, SFT, DPO, and reasoning (Chain-of-Thought), with a hardware-optimized inference engine using CUDA Graphs and static KV cache.
PyTorchLLM TrainingCUDADPOReasoningSynthetic Data
project
January 2025 – July 2025
EssayInsightsAI
Automated Essay Reviewer
Fine-tuned BERT variants for automated essay scoring and discourse segmentation, achieving 0.582 AUROC on Kaggle. Developed a Streamlit writing feedback assistant powered by LangChain.
BERTStreamlitOpenAI APILangChain
project
August 2024 – December 2024
ScratchNLP
NLP Algorithms from Scratch
A Python library with foundational NLP algorithms, preprocessing pipelines, tokenization techniques, and evaluation metrics from scratch supporting multiple datasets across several NLP tasks.
PythonPyTorchNLP
project
October 2022
Dysarthric Speech Clarity Improvement
Mel-spectrogram GAN Translation
Trained a CycleGAN on the TORGO speech dataset (10k samples) using Mel-spectrogram voice conversion to improve dysarthric speech clarity. Deployed as a web tool.
CycleGANFlaskJavaScriptAudio Processing
project
August 2022 – May 2023
Bachelor's Thesis Project
MGMT Methylation Prediction in Brain Tumors
Developed a deep learning and radiomics-based pipeline to predict MGMT promoter methylation in brain tumors. Ensembled SwinUNETR segmentation with texture features from PyRadiomics to achieve 0.609 AUROC.
SwinUNETRPyRadiomicsDeep LearningComputer Vision
project
May 2021
COVID-19 Chest scan Abnormalities Detection
Kaggle Bronze Medal Winner
Trained ChexNet and YOLO object detection models on TPUs to classify chest scans (6,300 images). Ensembled models to rank in top 23% on Kaggle.
ChexNetYOLOTPUMedical Imaging