class Trilochan:
location = "Nepal"
role = "AI/ML Researcher and Full-Stack Engineer"
spike = "Agentic memory systems, RAG pipelines, MCP tooling"
breadth = ["LLM fine-tuning", "Backend systems", "IoT and Edge AI"]
proof = {
"published_research": "doi.org/10.5281/zenodo.19784778",
"shipped_npm_package": "cms-mcp",
"live_products": ["AI portfolio platform", "autonomous job agent"],
}
def mission(self):
return "Build AI systems that survive contact with the real world"I learn whatever the problem requires and own the full pipeline — from data and training to deployment and monitoring. Currently CS undergrad at Kathmandu University, doing independent research on the side.
Every AI coding session starts blank — decisions, tradeoffs, and context from last week are gone. ContextForge is a persistent, queryable knowledge graph that gives AI coding assistants exactly the context they need across sessions, with adversarial-input defense built in.
| Result | Metric |
|---|---|
| Memory quality ranking | #1 of 6 systems benchmarked |
| Memory Integrity Score | 0.801 |
| Adversarial block rate | 90% (1% false positives in production mode) |
| Token savings vs. static context files | 93% |
| Benchmark suite | 990 tests passing |
📋 BibTeX
@software{sharma_2026_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}
}|
🛠️ cms-mcp · Published MCP server giving Claude programmatic control over any REST-based CMS. 32 tools, human approval gate with browser UI, policy engine, circuit breaker, audit logging, OpenAPI auto-discovery. 78 tests.
|
🧑💼 Job Agent Autonomous job application pipeline: scrapes listings, scores them with semantic search, generates ATS-optimized resumes via a DPO fine-tuned model, submits applications, and retrains nightly on real outcomes.
|
|
Multi-agent PDF tutor: upload any PDF, get a structured learning plan, cited teaching sessions, and evaluated practice questions. Response caching cuts API costs 60-80%. Fully local — no data leaves your machine.
|
Production-grade scene classification and image organization. MobileNetV2 transfer learning, ~86-87% accuracy, batch inference, auto folder sorting, ZIP export.
|
📁 More projects (classical ML, web)
| Project | What it does | Stack |
|---|---|---|
| AI Portfolio Platform | Dynamic portfolio with RAG chatbot, admin CMS, analytics. Live on Vercel. | Next.js 14, Supabase, Gemini |
| Product Pitch Recommender | Travel product recommendation with probability insights and bulk prediction | Scikit-learn, Streamlit |
| Customer Churn Prediction | End-to-end ML pipeline from preprocessing to Flask deployment | Scikit-learn, Flask |
| Student Placement Prediction | Placement prediction with full EDA and deployed Streamlit UI | Scikit-learn, Streamlit |
Languages
AI / ML — supervised and deep learning, transformers, fine-tuning (LoRA, DPO), evaluation and benchmarking
Agentic AI / RAG — multi-agent orchestration, vector search, MCP tool design, policy engines, prompt engineering
Web / Backend — REST APIs, auth (OAuth2, JWT, RLS), SSE, background jobs
Infra and Tools
IoT / Edge — ESP32, Raspberry Pi, Arduino, sensor pipelines, on-device inference (MobileNet), LoRa-based connectivity
| Area | What I'm doing |
|---|---|
| 🔬 Research | Agentic memory, adversarial defense, and RAG evaluation — follow-up work to ContextForge |
| 🌾 AgriTech | Architecting an offline-first precision agriculture platform for smallholder farmers in South Asia |
| 🧠 LLM internals | Implementing transformer architectures from scratch in PyTorch — attention, RoPE, MLA |
| ⚙️ Systems | Working through Designing Data-Intensive Applications, building distributed-systems intuition |
