000vectors indexed
HomeProjectsSkillsPlaySecurityAboutContact
Swaraj Chattaraj
Move to query.
Click to add a vector.

SwarajChattaraj

RAG, agents, things that ship
 RAG · LangGraph · pgvector · HNSW · DistilBERT · DeBERTa-v3 · Bucket4j · Supabase · Azure · Bun ·  RAG · LangGraph · pgvector · HNSW · DistilBERT · DeBERTa-v3 · Bucket4j · Supabase · Azure · Bun ·  RAG · LangGraph · pgvector · HNSW · DistilBERT · DeBERTa-v3 · Bucket4j · Supabase · Azure · Bun ·

Aria

before0.0s
after0.0s

-63% latency

Dristi

System 1 / System 2 router. DistilBERT answers in about 40ms, DeBERTa-v3 takes the hard cases.

Prosno

Chat with a GitHub repo. Hybrid retrieval, answers cite the code.
VECTOR, 384-D MINILM, PGVECTOR
FULL-TEXT, TS_RANK
MERGED, RRF

RAGPRO

Modular RAG pipeline engineered with LangChain, FAISS, and Hugging Face. Features 6 discrete swappable stages and sub-10ms similarity search.
STAGE 01 · INGESTION THROUGHPUT: 120 docs/sec

01 Load: Ingests raw repository source trees, markdown docs, and PDF specs through LangChain document loaders.

ComponentDirectoryLoader + PyPDF
InputRaw code & docs
OutputNormalized Docs

FUC
CLI

● BUN RUNTIME 
  
 
 
Skills
Pick a project to see what it used. Pick a skill to see where it showed up.
 
Full skills page →Play the games →
 
 
Projects
 
About
 
AI engineer building agentic RAG systems, routing infrastructure, and LLM-powered applications, from multi-agent pipelines to full-stack deployment. Currently researching clinical document understanding at Kalyani Government Engineering College, with a focus on making LLM outputs auditable and verifiable.
 
0projects built
0live apps
0open model on Hugging Face
 
NowAI Research Intern, Kalyani Government Engineering College, since Aug 2026. Multi-agent clinical document understanding.
EducationB.Tech, Artificial Intelligence, AGEMCO (MAKAUT)
GraduatingJuly 2027
FocusRAG, agent orchestration with LangGraph, production deployment
LanguagesEnglish, Bengali, Hindi
 
Skills
 
Play
Five arcade games, one for each project. Chase combos, beat your best, read the details after. Data is sample data. Mouse, touch and keyboard all work.
Pick a game above.
 
Say hello
 
 
 A+ POSTURE · AUDITABLE SYSTEMS 
Security
 

Production security policy, verifiable agentic audit loops, cryptographic hygiene, and interactive adversarial sandbox.

 
  
 
 
πŸ›‘οΈ
 Content-Security-Policy STRICT ENFORCED Whitelisted trusted CDNs, zero inline scripts 
 
 
πŸ”’
 Transport Layer HSTS PRELOAD (31536000s) TLS 1.3 enforced, auto HTTP-to-HTTPS redirect 
 
 
🚫
 MIME & Sniffing NOSNIFF · ZERO LEAK X-Content-Type-Options: nosniff active 
 
 
⚑
 Permissions Policy ZERO PERMISSIONS Camera, microphone & geolocation permanently blocked 
 
  
 
Verification & Defense Architecture
 
 
 
01
 
Clinical & Enterprise Verifier Loops
 

Researching clinical document understanding at Kalyani Government Engineering College. Designing multi-agent verifiers that audit generated outputs against exact source spans, ensuring critical healthcare summaries remain provably factual.

 
LangGraph · Span Matching · Automated Loopback
 
 
 
02
 
Human-in-the-Loop Sandboxing
 

In fuc-cli, autonomous tools are locked to realpath boundaries. All read operations execute immediately, while write operations (file edits, deletions, shell commands) require explicit cryptographic or operator signature before committing.

 
RealPath Jail · Staged Unified Diffs · Shell Locked
 
 
 
03
 
Out-of-Distribution Centroid Abstention
 

In Dristi, high-confidence queries run on DistilBERT in ~40ms, while anomalous inputs are evaluated against training centroid cosine distances with temperature calibration, preventing confident hallucinated predictions on adversarial payloads.

 
Cosine Distance · L-BFGS Scaling · Safe Abstain
 
 
 
04
 
Cryptographic Hygiene & Token Lifecycle
 

In Prosno, third-party GitHub OAuth tokens are encrypted at rest using AES-256-GCM. Inactive repositories transition through a timed expiration lifecycle (pending → indexing → ready → expired → hard-deleted).

 
AES-256-GCM · Scheduled TTL · Bucket4j Rate Limiting
 
 
 
  
 
Live Adversarial Input Sandbox
 
Test malicious payloads against the security verification filter:
 
 
   
 
     
 
 Enter a payload or choose a preset above to inspect security verification. 
 
 
 
 
Copied