╔══════════════════════════════════════════════════════════════════════════╗
║  PROCESS  : sahil.sharma                                                ║
║  ROLE     : Full-Stack AI Engineer · Systems Architect                   ║
║  STACK    : MERN · LangGraph · Reverse-RAG · Real-time ML Pipelines     ║
║  UPTIME   : B.Tech CSE Sem 5 @ JSS Academy, Noida                       ║
║  STATUS   : ▶  BUILDING SYSTEMS THAT THINK                              ║
╚══════════════════════════════════════════════════════════════════════════╝

I don't write features. I design intelligence layers — deterministic state machines that reason, hallucination firewalls that self-heal, and inference pipelines that scale. Every abstraction I write, I can trace back to the disk seek that executes it.

"The abstraction layer is a lease, not a permanent home. Own the metal beneath it."

ARCHITECTURE: enterprise orchestration + distributed ML inference
STATUS:       ██████████ active

A legal compliance engine that doesn't just flag dark patterns — it maps them to specific clauses in the DPDP Act, EU AI Act, and CCPA. Built on a five-layer pipeline: Puppeteer DOM extraction → YOLO + Open-CLIP visual deception detection → Transformer + XGBoost NLP ensemble → ChromaDB Legal RAG → Gemini cross-verification with automated multi-key failover.

Express Puppeteer YOLO Open-CLIP LangChain ChromaDB Gemini 2.5 XGBoost LightGBM

ARCHITECTURE: real-time streaming answer synthesis
STATUS:       ██████████ active

A Perplexity-class answer engine built from scratch. The non-obvious engineering here: a Socket.io streaming bridge that keeps token latency under 80ms end-to-end, a Redis JWT blacklist that prevents stale auth from hitting LLM endpoints, and Mistral Large used purely for deterministic title generation — not because it's cheaper, but because its output is more structurally consistent.

Node.js Socket.io LangChain Gemini 2.5 Flash Mistral Large Tavily AI Redis React 19

ARCHITECTURE: reverse-RAG stream interception
STATUS:       ████████░░ research

The core insight: instead of post-hoc fact-checking, intercept the LLM output stream before it reaches the client. LangGraph state machine routes live token emissions through MongoDB Atlas Vector Search for semantic factual validation. Corrections are injected mid-stream, not appended as disclaimers. The system heals its own output.

LangGraph MongoDB Atlas Vector Search Reverse-RAG Python FastAPI

ARCHITECTURE: parallel LangGraph state machines
STATUS:       ████████████ shipped

Performance duels between competing LLMs, orchestrated through LangGraph state machines that safely isolate and pipeline parallel model I/O. MongoDB aggregation handles win/loss tracking, category performance curves, and global leaderboard generation. Secure cookie architecture: HttpOnly, SameSite=None, Secure, with JWT refresh rotation.

LangGraph Cohere Gemini Mistral MongoDB TypeScript React

ARCHITECTURE: amazon-class storefront ecosystem
STATUS:       ██████░░░░ in progress

Not another CRUD store. Dual Buyer/Seller dashboards with RBAC + Google OAuth 2.0, a LangChain + LangGraph style recommendation engine that builds a "Style DNA" profile per user, and a GSAP + glassmorphism frontend that treats UI as a first-class product decision.

MERN LangChain LangGraph GSAP SCSS ImageKit JWT OAuth 2.0

ARCHITECTURE: on-device ML + adaptive UI
STATUS:       ████████████ shipped

Full biometric emotion classification running client-side — no round-trip to inference servers. MediaPipe classifies six emotional states at 30fps, Redis tracks session state and token invalidation arrays, and the UI adapts its layout and content in real-time. Zero-latency because the model never leaves the browser.

MediaPipe React Redis Node.js JWT

AI · ML · Orchestration

Backend · Systems

Frontend · UI

Data · Infrastructure

  

SAHIL(1)                    Developer Manual                    SAHIL(1)

NAME
    sahil — full-stack AI engineer, systems thinker

SYNOPSIS
    sahil [--build] [--research] [--obsess-over-fundamentals]

DESCRIPTION
    Builds production-grade AI systems as a B.Tech undergrad.
    Traces every abstraction to its machine-level origin.
    Ships systems most engineers won't touch for years.

OPTIONS
    --build           MERN + LangGraph + real-time ML pipelines
    --research        Reverse-RAG hallucination firewalls
    --fundamentals    B-Trees, OS memory layout, theory of computation
    --friends         Samarth, Meghanshu, Shubh
    --hardware        MacBook Air M4

PHILOSOPHY
    The abstraction layer is a lease, not a permanent home.
    Trace your queries to disk seeks.
    Map your variables to memory.

SEE ALSO
    github(1), langraph(1), mongodb-atlas-vector-search(1)

SAHIL                           2025                            SAHIL(1)

_{Systems don't lie. Code doesn't bluff. Ship or be shipped.}