Hi, I am Somay Kousis. I also go by Blink.
I am a student at ABV-IIITM Gwalior building in the overlap between GenAI, RAG, agentic AI, memory systems, and weird interfaces.
The thing I keep coming back to: software should not feel like a dead form. It should remember context, adapt to the person using it, and make collaboration feel less mechanical. That is the lane I am trying to get good at.
Right now I am focused on becoming internship-ready for AI/GenAI roles by shipping proof instead of just collecting buzzwords: working repos, demos, architecture notes, and systems that do more than wrap an API call.
- Memory for AI systems: long-term user memory, project memory, profile extraction, recall, and update flows.
- Agentic workflows: LangGraph-style routing, critique loops, background schedulers, and self-correcting pipelines.
- RAG that has personality: retrieval that separates factual context from tone/style context.
- Interfaces for AI tools: dashboards, cockpit views, state visualization, and products that feel human without becoming vague.
A stateful co-founder assistant that captures engineering logs, extracts persistent developer preferences, builds project dossiers, and runs self-correcting retrieval loops.
Signal: Python, FastAPI, LangGraph-style graph design, Supabase, Chroma, Groq/Llama, memory stores, background scheduling, structured logging, frontend dashboard.
A portfolio RAG system built on top of personal memory: identity, projects, writing, values, contradictions, current work, and tone.
Signal: Python, FastAPI, Jina embeddings, Supabase pgvector, prompt design, metadata-aware retrieval, factual/style context separation.
My interactive portfolio: part website, part environment, part proof that I care about how software feels.
Signal: Next.js, TypeScript, WebGL/particles, motion, frontend craft, AI-layer integration.
- Something: full-stack founder/investor platform idea with AI personas for critical and empathetic idea validation. Early stage.
- The Last Library: small RAG mystery project built while learning retrieval systems.
- SteelCareer: career platform/client-style product work.
- Customer Churn Prediction: ML project around churn risk and classification.
- House Price Prediction System: end-to-end ML pipeline with regression models.
AI / Backend
Python · FastAPI · LangGraph · LangChain · RAG · Chroma · Supabase · pgvector · Groq · Llama · Jina Embeddings
Frontend / Product
Next.js · React · TypeScript · Tailwind CSS · Three.js · Framer Motion
Tools / Infra
Git · Linux · Vercel · Postgres · Node.js · Express · MongoDB
- Shipping one strong project to completion instead of starting five interesting ones.
- Making AI systems evaluable: grounded retrieval, hallucination checks, memory quality, and failure modes.
- Writing clearer READMEs: demo, architecture, tradeoffs, setup, limitations.
- DSA consistency for internship interviews.
