"Building systems that reason, remember, and act in the physical and digital world."

Senior Applied Data Scientist by profession, open-source AI builder and tech enthusiast by obsession.

I'm deeply fascinated by the shift from passive next-token prediction to proactive, stateful multi-agent systems that can plan, coordinate, debate, and reliably interact with software and databases. My weekend builds focus on turning bleeding-edge ideas into runnable code — from cyclic deduction graphs in LangGraph to tool integration using Anthropic's Model Context Protocol (MCP), fine-tuning small models with LoRA/DPO, and building rigorous LLM-as-a-judge evaluation frameworks.

• Stateful Cyclic Graphs (LangGraph): Building multi-agent systems with independent memory checkpointers, parallel stateful nodes, dynamic routing, and human-in-the-loop validation.
• Model Context Protocol (MCP): Implementing standards-compliant JSON-RPC agent tool servers that decouple reasoning from data stores while enforcing role-based boundaries (RBAC) and database concurrency.
• Agentic Evaluation & Anti-Sycophancy: Designing automated evaluation harnesses (LLM-as-a-Judge, adversarial simulation panels, and benchmark validation) that test edge cases and prevent hallucination.
• Model Fine-Tuning & Distillation: Experimenting with parameter-efficient fine-tuning (LoRA, QLoRA, DPO, GRPO) to distill high-reasoning capabilities into fast, localized open weights.

Interactive multi-agent deduction game with independent stateful personas and forensic ground-truth validation.

• Concept: Every suspect is an autonomous, stateful LangGraph node with their own personality, secrets, and alibis. The user interrogates them as detective, unlocking physical/forensic clues in real time.
• Architecture: Plot Architect → Stateful Suspect Nodes (parallel memory checkpointers) → Clue Dropper Node (intent & keyword recognition) → Truth Keeper Node (indictment deduction evaluation).
• Playable Anywhere: Features both an interactive Terminal CLI (python play.py) and a sleek Web Arena (python app.py) with sub-second latency via Gemini 2.0 Flash and deterministic offline simulation.
• Stack: Python 3.12, LangGraph, Gemini 2.0 Flash, FastAPI, Pytest (18 passing tests).

flowchart LR
    Detective[Detective Question] --> Router{Suspect Router}
    Router --> SuspectA[The Butler: Alibi Defense]
    Router --> SuspectB[The Heiress: Secret Guard]
    Router --> SuspectC[The Physician: Forensic Trail]
    SuspectA & SuspectB & SuspectC --> ClueEngine[Clue Dropper Agent]
    ClueEngine --> Evidence[(Forensic Locker)]
    Detective -->|Formal Indictment| TruthKeeper[Truth Keeper Node]
    TruthKeeper --> Scorecard([Detective Scorecard: Sherlock Rank])

Autonomous residential operations coordinating tenants, owners, and management via Gemini and Model Context Protocol.

• Concept: Multi-persona agent pipeline (Tenant Agent, Owner Agent, Admin Agent, Supervisor FSM) communicating over MCP to manage apartment move logistics.
• Highlights: Concurrency-safe service elevator reservation using lazy-TTL mutex locks; digital gate pass generation with 6-digit OTP verification; human-in-the-loop approval checkpoints.
• Stack: Python 3.11+, FastAPI, Model Context Protocol (MCP), aiosqlite WAL, Pytest (28 tests passing).

Context-grounded technical interview simulator with realistic interviewer personas and dynamic pushback.

• Concept: Calibrates pushback against uploaded Job Descriptions and candidate resumes, probing distributed systems tradeoffs, sharding, and latency budgets without sycophantic praise.
• Stack: Next.js 15, LangGraph.js, Gemini 2.0 Flash, Tailwind CSS, Vitest.

Multi-target machine learning pipeline forecasting facility utilization across residential communities.

• Highlights: Feature engineering pipeline benchmarking LightGBM against historical baseline models with chronological validation and automated CI quality gates.
• Stack: Python, Scikit-Learn, LightGBM, Pandas, Pytest, GitHub Actions.

Always excited to chat about agentic architectures, autonomous systems, open-source AI tooling, or interesting research papers:

• LinkedIn: linkedin.com/in/prajwal-rao-d
• GitHub: @Riser01
• Email: [email protected]