class Trilochan:
name = "Parnish (Trilochan Sharma)"
location = "Nepal π³π΅"
role = "AI/ML Researcher & Full-Stack Engineer"
research = ["Agentic Memory Systems", "Adversarial AI Defense", "RAG Pipelines"]
focus = ["Agentic AI", "LLM Fine-tuning", "MCP", "RLHF"]
published = True # doi.org/10.5281/zenodo.19784778
mission = "Build impactful AI/ML solutions & contribute to cutting-edge research"
fun_fact = "I learn any tech needed to own the full pipeline β soup to nuts π"
def greet(self):
return "Let's build something that actually matters π"| π Problem Solver End-to-end across AI, ML, DL, NLP, Web, App & IoT |
π Published Researcher Peer-reviewed work on agentic AI memory systems |
π± Lifelong Learner Always upskilling in AI/ML, math & modern engineering |
β‘ Passion Emerging tech, complex problems, mastering new frameworks |
|
ContextForge: Agentic Memory for AI-Assisted Development Trilochan Sharma β Independent Researcher, 2026
|
π BibTeX Citation
@software{sharma_2025_contextforge,
author = {Sharma, Trilochan},
title = {ContextForge: Agentic Memory for AI-Assisted Development},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.19784778},
url = {https://doi.org/10.5281/zenodo.19784778}
}mindmap
root((TRILOCHAN ENGINE))
Research
Agentic Memory
Adversarial AI
RAG Systems
Benchmarking
AI/ML
Supervised Learning
Unsupervised Learning
Deep Learning
Transformers
RLHF
Backend
FastAPI
Node.js
Django
Flask
GraphQL
Frontend
React
Next.js
Tailwind
TypeScript
Data
PostgreSQL
MongoDB
Redis
Vector DB
SQLite
DevOps
Docker
Kubernetes
CI/CD
Monitoring
Cloud
AWS
Azure
GCP
Vercel
Linear Algebra β’ Calculus β’ Probability & Statistics β’ Discrete Mathematics β’ Graph Theory β’ Optimization β’ Mathematical Modeling
Core Skills: Supervised & Unsupervised Learning β’ Feature Engineering β’ Model Evaluation β’ Pipeline Design β’ Hyperparameter Tuning β’ Transfer Learning β’ Model Fine-tuning β’ DPO (Direct Preference Optimization)
Deep Learning: CNNs β’ RNNs β’ Transformers β’ GANs β’ LLM Applications β’ Transfer Learning β’ MobileNetV2 β’ Fine-tuning
NLP: Tokenization β’ Embeddings β’ Text Classification β’ Sentiment Analysis β’ Named Entity Recognition β’ Prompt Engineering
Core Skills: Multi-Agent Architecture β’ RAG Pipelines β’ Vector Search β’ LLM Orchestration β’ Prompt Engineering β’ Cosine Similarity β’ Embedding Pipelines β’ Response Caching β’ MCP (Model Context Protocol) β’ Tool Design β’ Policy Engines β’ Semantic Search β’ Circuit Breakers β’ Audit Logging β’ OpenAPI Auto-discovery β’ RLHF
Skills: Exploratory Data Analysis (EDA) β’ Statistical Analysis β’ Data Cleaning β’ Advanced Visualization β’ Feature Selection
Frontend: HTML β’ CSS β’ JavaScript β’ TypeScript β’ React β’ Next.js 14 (App Router) β’ shadcn/ui β’ Tailwind CSS
Backend: Node.js β’ Express.js β’ FastAPI β’ Django β’ Flask β’ REST APIs β’ GraphQL β’ SSE
Security & Auth: Swagger β’ OAuth2 β’ JWT β’ AES-256 Encryption β’ Supabase Auth β’ RLS
Databases: MongoDB β’ PostgreSQL β’ pgvector β’ MySQL β’ Redis β’ SQLite β’ Aerospike β’ Supabase
ORM / Data Access: Hibernate β’ JPA β’ Prisma β’ Drizzle
Platforms: AWS (S3 β’ EC2 β’ Lambda) β’ Azure β’ GCP β’ Vercel Β |Β Object Storage: AWS S3 β’ Aerospike β’ Supabase Storage
Raspberry Pi β’ Arduino β’ ESP32 β’ Sensor Integration β’ Data Pipelines β’ Edge Computing β’ NVIDIA Edge AI β’ Wireless Protocol Design
Containerization: Docker β’ Docker Compose β’ Kubernetes Β |Β CI/CD: GitHub Actions β’ Linux CLI β’ Bash Β |Β Testing: JUnit β’ Mockito β’ Zod Validation
|
π§βπΌ Job Agent Fully autonomous end-to-end job application pipeline. Scrapes LinkedIn, Indeed & Glassdoor, scores listings via semantic search, generates ATS-optimized resumes using a DPO fine-tuned model, and autonomously submits applications. Includes a nightly RL feedback loop that retrains on real outcomes (rejection, view, interview).
|
π οΈ cms-mcp Β· Published MCP server giving Claude programmatic control over any REST-based CMS (Supabase, Strapi, Payload). Features 32 MCP tools, human approval gate (browser UI + SSE), policy engine with 10 rule types, semantic search, circuit breaker, audit logging & OpenAPI auto-discovery. 78 tests.
|
|
π MeroStudySathy Intelligent multi-agent PDF tutor. Upload any PDF β structured learning plan, interactive teaching sessions with citations, follow-up chat & evaluated practice questions. Full RAG pipeline with 4 specialized agents. Response caching cuts API costs by 60β80%. Fully local β zero data leaves your machine.
|
Dynamic portfolio with RAG-based AI chatbot, admin CMS, analytics with CSV export & Supabase backend. Live on Vercel.
π LFFTT Full-stack web app with responsive design, state management & backend integration.
|
|
π ContextForge Β· Agentic memory system giving AI coding assistants persistent memory across sessions. Ranked #1 of 6 systems in memory quality (MIS=0.801). 90% adversarial block rate, 93% token savings, 990 benchmark tests, Ξ¦=79.7% composite safety index.
|
|
π¬ Scene Sorter Production-grade scene classification & smart image organization system. MobileNetV2 transfer learning with FastAPI backend + Next.js frontend. Supports single & batch inference, auto folder sorting, ZIP export. Achieved ~86β87% accuracy with optimized inference latency.
|
|
π― Intelligent Product Pitch Recommendation Advanced ML solution recommending optimal travel products. Single & bulk predictions, probability insights, CSV upload & interactive Streamlit UI.
|
π Student Placement Prediction Predicts placement based on IQ, CGPA, skills & internship experience. Full EDA, model building, evaluation & Streamlit deployment.
|
|
π Customer Churn Prediction End-to-end ML pipeline with preprocessing, feature engineering, modeling & Flask deployment.
|
Comprehensive EDA & visualization with statistical techniques and interactive charts.
|
| # | Area | What I'm doing |
|---|---|---|
| π | AI Research | Publishing work on agentic memory, adversarial defense & RAG systems |
| π¬ | Transformer Architectures | Exploring advanced LLM internals and architecture variants |
| π€ | Agentic AI Systems | Building production-grade autonomous pipelines with real-world automation |
| π§ | MCP Research | Tool design, policy engines, cross-LLM memory & orchestration |
| π | RL Feedback Loops | Self-improving AI pipelines that learn from real outcomes |
| π | Real-World Impact | Solving problems that actually matter |
"I believe in learning by doing. Every project is an opportunity to deepen understanding and create value. My mission is to build impactful AI/ML solutions and contribute to cutting-edge research and development."
β‘ Fun Fact: I enjoy learning any technology needed to solve real-world problems end-to-end β from data engineering to deployment, I care about the complete pipeline.
I'm always open to connecting on:
| π€ AI/ML Collaboration | π¬ LLM & MCP Discussions | π Knowledge Sharing | π Real-World Problem Solving |
|---|
If you find my projects useful β star a repo, share it, or drop feedback! Every β keeps the momentum going.
Made with β€οΈ by Parnish β Always learning, always building