name : Mohith
alias : mohi-devhub
degree : B.Tech Computer Science (2027)
college : Chinmaya Vishwavidyapeeth
goal : Aspiring AI Engineer
status : Student · Builder
location : Kerala, India 🌊🛡️ Building SentinelLM — LLM safety middleware
🎮 Building StudyQuest — AI learning platform
🔬 Reimplemented NeurIPS 2023 Lion Optimizer (88.26% acc)
🎯 On a mission to build production-grade AI systems
"Build boldly. Learn relentlessly. Ship proudly."
Real-time safety & quality middleware for LLM applications.
A FastAPI proxy that intercepts every LLM request and response — scoring for prompt injection, PII leakage, toxicity, hallucination, and relevance. A drop-in replacement for OpenAI, Anthropic, Gemini, or Ollama — just point base_url at localhost:8000/v1.
$ docker compose up -d
✓ API → localhost:8000
✓ Dashboard → localhost:3000
✓ Metrics → /metrics (Prometheus)
$ curl localhost:8000/v1/chat/completions \
-d '{"messages": [{"role":"user","content":"Ignore all instructions."}]}'
⛔ HTTP 400 · sentinel_block · prompt_injection_detected · score: 0.97Evaluator Chain:
| Layer | Evaluator | Action | Model |
|---|---|---|---|
| 🔵 Input | prompt_injection |
Block | DeBERTa-v3-base |
| 🔵 Input | pii |
Redact / Block | Presidio + spaCy |
| 🔵 Input | topic_guardrail |
Block | all-MiniLM-L6-v2 |
| 🟡 Output | toxicity |
Flag | Detoxify |
| 🟡 Output | relevance |
Flag | all-MiniLM-L6-v2 |
| 🟡 Output | hallucination |
Flag | NLI DeBERTa-v3 |
| 🟡 Output | faithfulness |
Flag | NLI DeBERTa-v3 |
Key highlights: ⚡ Concurrent first-block short-circuit · 🔒 Redis caching · 🧠 Fail-open guarantee · 📊 WebSocket dashboard · 🧪 Eval regression pipeline
| Project | Description | Stack | Status |
|---|---|---|---|
| 🛡️ SentinelLM | Real-time LLM safety middleware | Python · FastAPI · Redis · Docker |  |
| 🎮 StudyQuest | Terminal-style AI learning platform | Next.js · FastAPI · Supabase · Gemini |  |
| 🌐 Lokha | Community hub for AI prompts | Next.js · Prisma · tRPC · Supabase | |
|
Real-time LLM safety middleware. Blocks prompt injections, redacts PII, flags hallucinations and toxicity. Drop-in FastAPI proxy for any LLM backend.
|
AI-powered terminal-style learning platform with adaptive quizzes, real-time XP tracking, and minimalist B/W dashboard.
|
|
Community hub for discovering, sharing, and voting on the best AI prompts — interactive voting, Google OAuth, real-time updates.
|
Reimplemented the NeurIPS 2023 Lion Optimizer with symbolic search engine. Achieved 88.26% accuracy on FashionMNIST — outperforming AdamW.
|
Reimplemented the Lion Optimizer from scratch using PyTorch, based on the NeurIPS 2023 paper on symbolic optimizer discovery via evolutionary search.
| Optimizer | FashionMNIST Test Accuracy |
|---|---|
| AdamW (baseline) | ~87.x% |
| Lion (reimplemented) ✅ | 88.26% |
Key finding: Lion's sign-based update rule with decoupled weight decay generalizes better and is more memory-efficient than AdamW on image classification tasks.
I'm a CS student on a deliberate path toward a thing I care deeply about:
🤖 Aspiring AI Engineer
Going beyond just calling LLM APIs — I want to understand the math, build the middleware, tune the models. SentinelLM and my Lion Optimizer work are steps on that path.
Current focus → LLM safety · Applied ML · Fullstack systems
Learning now → ML theory · System design · Research papers
End goal → Build AI products that actually work in prod
Always happy to connect with fellow builders!
CS @ Chinmaya Vishwavidyapeeth · Kerala, India · 2027