Skip to content

AI work: Finfluencer agent, AWS AI League, DeepRacer

Personal Financial Intelligence Agent ("Finfluencer AI")

Google Cloud Agentic AI Day (Hack2Skill), July 2025. Shortlisted in the Top 700 teams of 57,000+ developers.

A personal financial advisor that reads your real money data through Fi Money's MCP server: net worth, credit report, EPF, mutual funds, stocks and bank transactions. You give it a goal like "open a dance studio" or "retire by 50" and it turns that into a plan. One orchestrator agent routes to specialist agents (MoneyMind for cash flow, Lakshya AI for goal plans) and asks for missing inputs before it plans. Built with Gemini, Google's Agent Development Kit and Vertex AI, with a Streamlit front end.

Architecture: question → orchestrator → Fi MCP tools → specialist agents → Gemini → plan.

What is MCP?

The Model Context Protocol: an open standard for connecting LLM applications to tools and data. A server exposes tools (functions the model can call), resources (data it can read) and prompts; a client (the agent host) discovers them and lets the model call them. It replaces one-off integrations per model and tool with one protocol, "USB-C for AI tools". Fi Money exposed a user's financial data as an MCP server; our agent was the client. Hackathon teams used a mock server with dummy data and a dummy login, so no real user data.

Why multiple agents instead of one big prompt?

Each specialist has a smaller, focused prompt and tool set, so it is more reliable and easier to test; the orchestrator only routes and asks for missing inputs. Cost: more LLM calls and latency, and routing errors. One agent is fine until its prompt gets crowded with unrelated instructions.

How do you stop an agent giving bad financial advice?

Guardrails: ground every number in tool data (no invented figures), ask for missing inputs instead of assuming, show the data behind each recommendation, limit to education/planning (not regulated advice) with a disclaimer, validate outputs (e.g. totals add up), and keep a human decision at the end.

What is context engineering?

Deciding what goes into the model's context window for each call: instructions, the relevant data (only what this step needs), tool results, conversation memory, examples. More context is not better: irrelevant context distracts and costs tokens. Summarise history, retrieve only relevant data, keep tool outputs compact.

What would you do with more time?

Evaluations (a set of test questions with expected answers, run on every prompt change), caching tool results, streaming responses, and tracing each agent step for debugging.

AWS AI League: city lead & SME, Mumbai tech center (2025–2026)

AWS's gamified agentic-AI competition. Participants built multi-agent systems on Amazon Bedrock (Agents, Lambda) to learn prompt engineering and agent design.

  • What I did: ran the site program (Fill in: participants, sessions, results): onboarding, workshops, office hours as the subject-matter expert, unblocking teams, coordination with AWS and internal sponsors.
  • Leadership angle: leading without authority, explaining agent design to engineers who are not ML specialists, keeping momentum over weeks.
What is a Bedrock Agent?

A managed agent on AWS: you give it a foundation model, instructions, action groups (tools, usually Lambda functions described with an OpenAPI schema) and optionally knowledge bases (RAG over your documents). Bedrock runs the reason → call tool → observe loop. Multi-agent collaboration lets a supervisor agent delegate to sub-agents.

What is RAG?

Retrieval-Augmented Generation: embed documents into vectors, store them in a vector index; at question time, retrieve the most similar chunks and put them in the prompt so the model answers from them. Reduces hallucination and lets the model use private, current data without fine-tuning.

DeepRacer

  • 2023: 2nd place, led a team of 6. Nobody on the team could tell why the car made certain decisions, so I built a tool that drew what our reward function was actually rewarding on the track. Seeing it made the fixes obvious; that is how we climbed to 2nd.
  • 2024: co-led the competition for 700+ participants.
How does DeepRacer learn?

Reinforcement learning: the car (agent) observes the camera image, picks an action (steering, speed), and gets a reward from a function you write (e.g. stay near the centre line, go fast on straights). Training (PPO) adjusts the policy to maximise total reward. The skill is designing a reward function that rewards what you want, not something the car can exploit.

The common thread

"I make complex systems visible": the DeepRacer reward visualiser, Prayog's dashboards and replay, Galois explaining the maths. Use it as the theme if they ask "what kind of engineer are you?"