Soham Chaudhari, AI Engineer

AI Engineer

I build agentic systems, retrieval pipelines and multimodal vision, and take them from notebook to product.

Now2026

AI Engineer at FLIP

A platform that helps people find card offers, earn more rewards and redeem points well.

  • Agentic AI
  • LLM Orchestration
  • RAG
  • Voice AI
  • Fintech AI
LangChainLangGraphRAGAgentic AIPrompt EngineeringFunction CallingLLM EvaluationHugging FacePyTorchTensorFlowKerasscikit-learnTransformersCLIPComputer VisionOptuna
GPTsai/mlAI Agentsai/mlStock AIai/mlFace Detectionai/mlMNIST CNNai/mlRegressNCompareai/mlVISION AIai/mlCold email generatorai/mlMOSAICai/mlAutogen Agentsai/mlRAG Doc-Chatai/mlAutogen Multi-Agent Systemai/mlLanguage-Detectorai/mlWeb scrapperai/mlGemini Cloneai/mlNeural Networksai/mlODINai/mlWhatsApp Message Automationai/mlMedicXai/mlLangGraph Agentsai/mlSentence Transformers Deployai/mlMail Sendai/ml
[01]

About

Portrait of Soham Chaudhari
Mumbai, India
At a glance

AI Engineer with applied experience Building production AI Systems across fintech, legal-tech and consumer platforms.

I care about the unglamorous parts too: evals, latency, cost.

Now
AI Engineer at FLIP
Based in
Mumbai, India
Studied
B.E. Computer Engineering, 2026
Resume GitHub
[02]

Selected work

06 / 30 systems

6 systems that see, read and act.

02.1 / AI / MLvision

MOSAIC

MOSAIC (Multimodal Orchestration for Synthesis, Analysis & Intelligent Comprehension) - AI-powered video analysis platform featuring FastAPI backend, React frontend, and MCP server. Process videos, extract clips, ask questions about content, and perform multimodal search using vision models and transcription, Even creates clips from a video!

  • MCP
  • Vision
  • Speech
  • FastAPI
Case study Source
02.2 / AI / MLlanguage

GPTs

Decoder-only GPT language models built from scratch in PyTorch: token and position embeddings, multi-head causal self-attention, residual transformer blocks and an LM head (~10M params), trained with AdamW for character-level text generation. A home for small language models trained from the ground up.

  • LLM
Case study Source
02.3 / AI / MLagents

Stock AI

An LLM stock-analysis agent served as a FastAPI service. A Groq-hosted gpt-oss-120b agent calls tools for ticker search, live prices, company metrics, income statements and Yahoo Finance screeners, and keeps per-conversation memory. Also exposes the same tools as an MCP server, with API keys, rate limiting and caching.

  • MCP
  • LLM
  • FastAPI
Case study SourceTry it live
02.4 / AI / MLvision

VISION AI

An AI-powered solution for efficient video and image analysis, leveraging advanced machine learning models to enhance visual data processing and superscaling solutions.

  • Vision
  • ML
Case study SourceLive demo
https://hack-frontend-15p7qmfx4-sohamchaudhari2004s-projects.vercel.app/
VISION AI screenshot
02.5 / AI / MLvision

Face Detection

Real-time face detection that runs entirely in the browser. The YuNet ONNX model finds faces and five facial landmarks in photos or a live webcam feed, with no server and no image ever uploaded. Computer vision with ONNX Runtime Web.

  • Vision
Case study Try it live
02.6 / AI / MLml

MNIST CNN

A convolutional neural network trained in PyTorch on MNIST handwritten digits, exported to ONNX and served in the browser. Draw a digit and the model classifies it live, showing its confidence for every class.

  • Deep Learning
Case study Try it live
[03]

Ex plore

index of experiments
Hover to preview · click to open repo
[04]

Experience

03 roles

Where I've shipped.

  1. 2026 - PresentCurrent

    Remote

    AI Engineer at FLIP

    Building the AI layer of a fintech platform that helps people get more out of their credit cards: find the right offer, earn the most on every purchase, and redeem points for real value.

    • Built a multi-agent AI assistant that answers plain-language questions about credit card and bank offers.
    • Added a voice agent on top of it for hands-free, spoken conversations about offers.
    • Developed a rewards engine that recommends the highest-earning card for any purchase.
    • Building a points redemption experience that finds the best-value way to use points across 50+ cards.
    • Agentic AI
    • LLMs
    • Voice AI
    • Recommendations
    • Fintech
    • 01 / AI chatbot

      Offer Search Agent

      Ask "any offers on Myntra?" and get the relevant card and bank offers back, ranked for you.

      • Understands natural-language questions about brands, categories, banks and cards.
      • Finds, filters and ranks the offers that actually apply to the user.
      • Remembers the conversation, so follow-up questions just work.
      • Agentic AI
      • LLMs
      • Conversational AI
    • 02 / Voice

      Voice Agent

      The offer assistant, hands free. Ask out loud, hear the best offer back.

      • Full spoken loop: the user talks, the agent reasons, and it answers in a natural voice.
      • Built for quick, on-the-go questions while shopping.
      • Voice AI
      • Speech-to-text
      • Text-to-speech
    • 03 / Earn

      Rewards Calculator

      Tells you which card earns the most on a purchase before you pay.

      • Enter a brand and an amount; the system works out what kind of spend it is.
      • Compares the rewards every card would earn on it.
      • Recommends the highest-earning card, with the reward value.
      • Recommendations
      • Rewards logic
    • 04 / RedeemIn progress

      Points Redemption

      Finds the best way to spend accumulated points across 50+ cards.

      • Pick a card and a points balance to see every option: airline and hotel partners, cash and products.
      • Each option shows points needed, value, expiry and transfer time.
      • Example: turning card points into airline miles for a flight booking.
      • Optimisation
      • Product UX

  2. 2025 - 2026

    Mumbai

    AI Developer Intern at UNLAWC

    Legal-tech AI agents for document analysis and legal workflows.

    • Shipped end-to-end legal AI agents that cut manual review effort by about 50%.
    • Architected RAG pipelines on Mistral, LangChain and Pinecone with reranking and self-correction, lifting retrieval accuracy by 60%.
    • Raised contextual accuracy by 31% through structured outputs and evaluation-driven prompt iteration.
    • Reduced inference cost by about 80% with optimised retrieval and model routing.
    • Mistral
    • LangChain
    • Pinecone
    • RAG
    • Evals
    • 01 / Agents

      Legal AI Agents

      End-to-end agents for legal document analysis and review workflows.

      • Took over the repetitive parts of document review, cutting manual effort by about 50%.
      • Structured outputs and evaluation-driven prompt iteration raised contextual accuracy by 31%.
      • Mistral
      • LangChain
      • Structured outputs
      • Evals
    • 02 / RAG

      Legal RAG Pipeline

      Retrieval over legal documents that stays accurate and cheap at scale.

      • Mistral, LangChain and Pinecone with reranking and a self-correction step.
      • Lifted retrieval accuracy by 60%.
      • Optimised retrieval and model routing cut inference cost by about 80%.
      • Pinecone
      • Reranking
      • Model routing

  3. 2024 - 2025

    Mumbai

    SDE Intern at SR Counselling India

    ML recommendations and the backend that serves them.

    • Built a production recommendation engine handling 10K+ daily interactions at 95% accuracy, lifting engagement by 35%.
    • Built 15+ REST APIs on Node.js and Express serving 5K+ concurrent requests.
    • Designed an AI-powered visa training system that cut training costs by about 90%.
    • Shipped React admin dashboards with 60% faster page loads.
    • Python
    • Node.js
    • Express
    • React
    • Flutter
    • 01 / ML

      Recommendation Engine

      Production ML recommendations for the counselling platform.

      • Handles 10K+ daily interactions at 95% accuracy.
      • Lifted user engagement by 35%.
      • Python
      • ML
    • 02 / AI

      AI Visa Training System

      AI-powered visa training that replaced most of the manual effort.

      • Cut training costs by about 90%.
      • Python
      • AI
    • 03 / Backend

      Platform APIs & Dashboards

      The services and admin tools behind the platform.

      • 15+ REST APIs on Node.js and Express serving 5K+ concurrent requests.
      • React admin dashboards with 60% faster page loads.
      • Node.js
      • Express
      • React

Education

B.E. Computer Engineering

Vivekanand Education Society's Institute of Technology, Mumbai

2022 - 2026 · CGPA 8.39 / 10

Certifications

  • Oracle · Oracle Cloud Infrastructure 2025 AI Foundations Associate
  • NVIDIA · Fundamentals of Deep Learning
  • BCG · GenAI Job Simulation
  • Postman · API Bootcamp
[05]

The stack

40 params · 6 layers

A forward pass through my toolkit.

Languages feed the network, frameworks do the heavy lifting in the hidden layers, and what I care about sits at the output as the objective.

L00 · input05 params

Programming Languages

  • Python
  • TypeScript
  • JavaScript
  • Java
  • SQL
L01 · hidden08 params

GenAI & LLMs

  • LangChain
  • LangGraph
  • RAG
  • Agentic AI
  • Prompt Engineering
  • Function Calling
  • LLM Evaluation
  • Hugging Face
L02 · hidden08 params

ML & Deep Learning

  • PyTorch
  • TensorFlow
  • Keras
  • scikit-learn
  • Transformers
  • CLIP
  • Computer Vision
  • Optuna
L03 · hidden06 params

Vector & Databases

  • FAISS
  • Pinecone
  • ChromaDB
  • PostgreSQL
  • MongoDB
  • Redis
L04 · hidden06 params

Backend & APIs

  • FastAPI
  • Flask
  • Node.js
  • Express
  • React
  • Next.js
L05 · hidden07 params

MLOps & Cloud

  • AWS
  • Docker
  • Kubernetes
  • MLflow
  • GitHub Actions
  • Terraform
  • Git
L06 · output04 params

Objective function

  • ↳Generative AI
  • ↳Agentic AI
  • ↳Multimodal AI
  • ↳Deep Learning
[06]

Achievements

hackathons

Tested against 1,700+ builders under a clock.

01 · winner2026
#01st

Syrus

@Syrus'26 Hackathon

Won Syrus'26 among 150+ teams with Common Brain, a RAG assistant that unifies scattered organisational documents into one queryable, source-grounded knowledge layer with role-aware retrieval.

#2
#1
#3
02 · podium2025
#02nd

IIT Ropar

@Medino'sXAdvitiya'25

Runner-up at Medino'sXAdvitiya'25, hosted by IIT Ropar, among 200+ teams. Built a symptom analyser chatbot and an OCR-based prescription reader.

#2
#1
#3
03 · leaderboard2024
top10

IIIT Dharwad

@Hack2Future 2024

Top 10 at Hack2Future among 1,700+ participants with Vision AI: real-time object detection, scene recognition and automated insights for video and images.

1,700+ participantstop 0.6%
[07]

Elsewhere

05 channels

Find me on the network.

Based in Mumbai, India (IST, UTC+5:30). Always happy to talk agents, retrieval and vision. Pick whichever channel suits you.

[01] Email
sohamrc08@gmail.com

Roles, projects, collaborations. The fastest way to reach me.

[02] LinkedIn

LinkedIn

in/sohamchaudhari2004

Experience, recommendations and the career timeline.

[03] GitHub

GitHub

@SohamChaudhari2004

Every repo on this page, and the experiments that didn't make it.

[04] X

X

@sohamrchaudhari

Build logs, papers I'm reading, half-formed ideas.

[05] Medium

Medium

@sohamrc08

Long-form write-ups on agents, RAG and vision.

[08]

Terminal.

Let's build something intelligent.

Start a conversation, grab the resume, or prompt the terminal. Agents, RAG, vision, or all three at once.

Resume
soham@latent-space: ~ · zsh
soham-os 2026.01 (latent-space) · tty1
type `help` to list commands, or tap a suggestion below.
SC/ latent-spacev2026.01

booting soham chaudhari

000
  • > init runtime ............................ ok
  • > load weights soham.safetensors [30 repos]
  • > tokenizer 40 tools · 6 layers
  • > agents autogen · langgraph · online
  • > evals #1 · #2 · top 10
  • > ready.