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, MCP-style tool orchestration, 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 building Something, a verified discovery layer where founders and investors find each other on proof, not luck. Still just a developer with vibes underneath it though, shipping working repos and real architecture instead of collecting buzzwords.
- Agentic AI systems: LangGraph routing, multi-agent debates, critique loops, background schedulers, and self-correcting pipelines.
- Memory for AI systems: long-term user memory, project memory, profile extraction, recall, and update flows.
- RAG that has personality: retrieval that separates factual context from tone/style context.
- MCP-style tool use: connecting models to tools, repositories, documents, and product workflows.
- Interfaces for AI tools: dashboards, cockpit views, state visualization, and products that feel human without becoming vague.
A verified discovery layer for founders, investors, and the ideas worth funding. Two entry paths gate visibility on proof of work instead of network access, matching runs bidirectional and reverse, and an agentic review pipeline sits underneath it as infrastructure, not the pitch.
An AI exploration engine for messy questions: it stages issues as a courtroom-style multi-agent debate, runs perspectives in parallel, lets a judiciary reason over public statements, and supports human-in-the-loop continuation or final conclusions.
A research companion built on a CockroachDB memory substrate with bi-temporal fact versioning and an ADD/UPDATE/INVALIDATE/NOOP consolidation write path, not a vector store bolted to a chatbot. Survived a live database crash mid-conversation with 25 out of 25 concurrent writes intact under SERIALIZABLE retries, where a flat file lost 24. Retrieval holds around 15ms at 10K notes on a C-SPANN ANN index.
A stateful co-founder assistant that captures engineering logs, extracts persistent developer preferences, builds project dossiers, and runs self-correcting retrieval loops.
A portfolio RAG system built on top of personal memory: identity, projects, writing, values, contradictions, current work, and tone.
My interactive portfolio: part website, part environment, part proof that I care about how software feels.
Smaller Experiments (click to expand)
- 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
Frontend / Product
Tools / Infra
- 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, mostly so it stops being the thing that trips me up in interviews.
