Skip to content
Name
Smit FalduAI / ML Engineer
Focus
Agentic Systems
Based in
Ahmedabad, IN
Local time
IST

Smit Faldu — AI / ML Engineer. I build systems that reason, retrieve, and act.

AI/ML engineer working on generative and agentic systems — multi-agent frameworks, retrieval-augmented pipelines and autonomous workflows, built with LangChain, LangGraph and AutoGen, and shipped as FastAPI services on GCP.

Currently

Building agentic systems and shipping them as services. Most recently AI/ML at VertxAI.

§01Selected Work

04 systems · 2025

Systems, notdemos.

Each of these was designed, built and deployed end to end — agent orchestration, retrieval, API surface and infrastructure.

01Featured

Mock Pitch Voice Agent

A real-time voice agent that runs investor Q&A sessions, so founders can rehearse the pitch before it counts.

Problem

Founders get one shot at an investor conversation and almost no way to rehearse the unscripted half of it — the questions. The system simulates that pressure: it listens, interrogates, adapts, and reports back.

Approach

  1. Agentic conversation flow built on LangChain and LangGraph, handling context-aware questioning, evaluation and adaptive branching as the pitch develops.

  2. Speech-to-text via Whisper and text-to-speech via ElevenLabs, wired into a continuous real-time loop rather than a turn-based transcript.

  3. FastAPI backend with WebSocket streaming to keep round-trip latency low enough for the exchange to feel conversational.

  4. Conversation data persisted and analysed in MongoDB, turning each session into performance tracking and feedback signal.

  5. Containerised with Docker and deployed on GCP Cloud Run, with a modular architecture that keeps new capabilities additive.

Role
Agentic AI · Voice Interface
Stack
  • Python
  • FastAPI
  • LangChain
  • LangGraph
  • WebSockets
  • Whisper
  • ElevenLabs TTS
  • MongoDB
  • Docker
  • GCP
Deployed
GCP Cloud Run
Year
2025
02

Founder–Investor Matching Engine

A semantic matching system that ranks the investors most likely to fund a given founder, from embeddings rather than keywords.

Problem

Founder–investor fit is a semantic problem wearing a search-box disguise. Keyword filters miss thesis alignment; the engine reasons over profiles instead of matching strings.

Approach

  1. LangChain and LangGraph pipelines covering query understanding, ranking logic and multi-agent reasoning to improve match quality beyond a single similarity pass.

  2. FAISS vector index retrieving the top 100 candidate investors per founder profile from contextual embeddings.

  3. FastAPI backend with WebSocket streaming, keeping the interaction responsive as results resolve.

  4. Match results stored in MongoDB so investor–founder interactions can be tracked, filtered and analysed over time.

  5. Containerised with Docker and deployed on GCP Cloud Run for production-ready scale.

Role
Retrieval · Multi-Agent Reasoning
Stack
  • Python
  • LangChain
  • LangGraph
  • FAISS
  • MongoDB
  • Docker
  • GCP
Deployed
GCP Cloud Run
Year
2025
03

ArXiv Research Agent

A multi-agent research assistant that retrieves, summarises and answers questions across ArXiv papers.

Problem

Keeping current with a fast-moving field is a retrieval problem and a comprehension problem at once. Splitting those across specialised agents produces better answers than asking one model to do both.

Approach

  1. Multi-agent architecture in AutoGen, with separate agents collaborating on retrieval and on summarisation.

  2. LLM-driven paper understanding, distilled into concise outputs written for a reader rather than a benchmark.

  3. Interactive Streamlit interface producing structured results in JSON and Markdown for downstream use.

  4. Containerised with Docker and published on Hugging Face Spaces for public access and reproducibility.

Role
Multi-Agent Research Systems
Stack
  • Python
  • AutoGen
  • Hugging Face Transformers
  • Streamlit
  • Docker
Deployed
Hugging Face Spaces
Year
2025
04

DataAnalyzerAI

Upload a CSV or JSON file, ask a question in plain language, and get back executed analysis with charts.

Problem

The distance between having data and understanding it is usually a Python session. This collapses that into a sentence.

Approach

  1. AutoGen multi-agent system that accepts CSV/JSON uploads and interprets natural-language queries against them.

  2. Agents generate, execute and return Python code, producing analysis and graph visualisations without user intervention.

  3. End-to-end exploration tool with reproducible results, packaged through Dockerised deployment.

Role
Agentic Data Analysis
Stack
  • Python
  • AutoGen
  • LLM
  • Streamlit
  • Docker
Deployed
Hugging Face Spaces
Year
2025

End of catalogue

More on GitHub

§02Capabilities

7 groups

The working set.

Grouped by the part of the system they belong to, not by how impressive they look in a list.

Agentic Systems

Orchestration, state and control flow across cooperating agents.

  • LangGraph
  • AutoGen
  • LangChain
  • Multi-Agent Architectures
  • Autonomous Workflows
  • Tool Use & Routing

Generative AI & LLMs

Model integration, conversation design and language understanding.

  • Large Language Models
  • Hugging Face Transformers
  • Conversational AI
  • Natural Language Processing
  • Whisper (STT)
  • ElevenLabs (TTS)

Retrieval & Data

Getting the right context in front of the model, quickly.

  • RAG Pipelines
  • FAISS
  • Embeddings
  • MongoDB
  • SQL
  • Pandas
  • NumPy

ML Foundations

The layer underneath the abstractions.

  • PyTorch
  • TensorFlow
  • Scikit-learn

Backend & Services

Shipping models as systems other software can call.

  • Python
  • FastAPI
  • WebSockets
  • REST APIs
  • Playwright

Infrastructure

Reproducible builds, scalable deploys, fast iteration.

  • Docker
  • GCP Cloud Run
  • Hugging Face Spaces
  • CI/CD

Interfaces

Enough front-end to make the system usable.

  • Streamlit
  • JavaScript
  • React
  • Tailwind CSS

Ways of working

  • Problem-Solving
  • Research & Experimentation
  • Rapid Prototyping
  • Critical Thinking
  • Collaboration
  • Adaptability

§03Trajectory

01 role · 02 qualifications

Where the work has been.

Focused on next-generation intelligent systems that combine reasoning, retrieval, and human–AI collaboration.

Practice
AI/ML engineer specialising in Generative AI, LLMs and Agentic AI systems.
Span
2021 — Present
Base
Ahmedabad, IN

Experience

01

Mar 2025 — Aug 2025

AI/ML Intern

VertxAIRemote · India

  • Contributed to Generative AI and Agentic AI pipelines, focused on LLM integration and retrieval-augmented generation.

  • Designed and implemented LangChain- and LangGraph-based workflows enabling multi-agent reasoning and task automation.

  • Built and deployed FastAPI backend services for AI-powered applications, with scalability and performance as constraints.

  • Optimised FAISS vector retrieval and embeddings pipelines, improving accuracy and latency in production workloads.

  • Worked with senior engineers to deploy on GCP (Cloud Run, Docker) and integrate CI/CD practices for faster experimentation.

Education

02

  • May 2024 — Present

    B.Tech, Computer Engineering

    SAL Institute of Technology & Engineering ResearchAhmedabad, Gujarat

  • Aug 2021 — May 2024

    Diploma, Computer Engineering

    Government Polytechnic, AhmedabadAhmedabad, Gujarat

§04Contact

Have somethingworth building?

Email

smitfaldu2006@gmail.com

Fastest way to reach me. Happy to talk about agentic architectures, retrieval pipelines, or anything adjacent.

GitHub
smit-faldu
LinkedIn
smit-faldu
Phone
+91 8140950661
Location
Ahmedabad, Gujarat, India