Vinayak Ajith
Agents, retrieval, and eventually robots.
I'm an AI/ML engineer. I build agents that work on a company's real data (its databases, email and documents) and answer with numbers you can check.
Now
OptiFlow, at Dellizo
OptiFlow is an AI analyst that runs on a company's own servers. You connect the databases, email and documents once, then ask questions in plain English. It writes the queries, reads the relevant threads, and answers with numbers that trace back to the query that produced them. When it doesn't know, it says so.
I built it end to end: the agent architecture, the email research subagent, the retrieval layer and the front end.
Focus right now: agent memory, hybrid retrieval and reranking, and OS internals for matter-os.
Work
-
matter-osearly
An experimental AI-native operating system on a custom RISC-V kernel.
RISC-V · kernel · agents
-
mnemo
A personal AI companion with long-term memory. It keeps facts and notable episodes across conversations and ranks them by relevance, recency and importance.
Python · pgvector · Mem0 · Claude
-
autonomous-research-agent
Researches a topic end to end: searches the web, reads PDFs, keeps its own state, and writes a structured summary. Runs on a local model.
LangGraph · Tavily · Llama 3
-
Hierarchical emotion classifierexperiment
130 emotions arranged as a feelings wheel, three levels deep. Masking keeps each prediction consistent from broad to fine, with a consistency loss and focal loss for the rare classes.
PyTorch · RoBERTa
Client work: admissions platform for Fluence International company site for KK Trading Co.
Path
-
Class 7
First program
A materials calculator for my father's truss work.
-
2021–2025
SRM Institute of Science and Technology, Chennai
B.Tech, Computer Science and Engineering.
-
Aug 2025
Joined Dellizo intern
The product wing of Ecosoft Solutions.
-
Feb 2026 – now
AI/ML Engineer, Dellizo
Building OptiFlow end to end.
Teachers
Three people whose free lectures taught me the fundamentals. Credit where it's due.
Stack
- Agents
- LangGraph state graphs, sub-agents, tool use and MCP, sandboxed code execution, Temporal and Celery for long-running work, Langfuse tracing
- Memory
- Long-term agent memory with consolidation (“dreaming”), episodic and semantic stores, Mem0, pgvector
- Retrieval
- Hybrid search (BM25 and vectors, reciprocal rank fusion), reranking, OCR and document parsing
- Models
- Claude, OpenAI, Gemini and open-weight models behind one routing layer; vLLM and Ollama, on rented RunPod GPUs
- Training
- PyTorch, Hugging Face Transformers, fine-tuning (RoBERTa, FLAN-T5), sentence-transformers
- Coding
- Lovable, then Cursor and Antigravity; now Claude Code and Codex, every day
- Product
- Python, TypeScript, SQL; FastAPI, PostgreSQL, Redis; React, ECharts; Docker


