Jay Rathod

AI Engineer · Mumbai

I build working systems where AI, software and hardware meet.

3+ years shipping LLM products and Python backends: chatbots with real memory, an invoice pipeline a business relies on, and a foot-pressure mat I took from bare sensors to a finished PDF report.

Signal chain · things I've built at every layer
  1. Sense Resistive sensor matrix
  2. Firmware ESP32 · C++ · multiplexers
  3. Backend FastAPI · Celery · PostgreSQL
  4. Model GPT · Gemini · DeepSeek · Qwen
  5. Output Heatmaps · PDF · Excel

01 · Selected work

Systems in production, beta and on the bench

01 Applied AI · Document processing

Purchase Invoice Extraction

Scanned purchase invoices in, validated Excel reports out. Built alone, deployed for a private firm.

Deployed, in client useInnovizia TechnologiesSole engineer: design, build, deployment

  1. 01 · Upload Batch of up to 10
  2. 02 · OCR Qwen3-VL 8B
  3. 03 · Structure GPT-4.1
  4. 04 · Validate Totals · GST · discounts
  5. 05 · Review Human in the loop
  6. 06 · Output Excel report
~5 min
for a batch of 10 invoices
50–100 min
estimated manual entry for the same batch
Read the case study

02 Embedded · Signal processing · Applied AI

FSP: Foot Pressure Mat

A low-cost foot-health screening product: sensor mat, ESP32 firmware, pressure heatmaps, guarded AI analysis and a PDF report.

Working prototypeInnovizia TechnologiesOwned end-to-end technical development

  1. 01 · Sense Resistive matrix
  2. 02 · Scan ESP32 + multiplexers
  3. 03 · Clean Normalise + de-ghost
  4. 04 · Map Foot regions
  5. 05 · Decide Rules engine
  6. 06 · Explain LLM + foot images
  7. 07 · Report PDF
300
people in the real test dataset
Concept → PCB
from first idea to a manufactured, assembled board
Read the case study

03 Generative AI · Conversational memory

My Best Self (MBS)

The AI pipeline behind a chatbot platform with configurable bots: memory, retrieval, model fallback and overnight jobs.

Beta on Google PlayInnovizia TechnologiesOwn the core AI pipeline

  1. 01 · Input User message
  2. 02 · Recall Short-term memory
  3. 03 · Retrieve Long-term + RAG
  4. 04 · Generate GPT-4o mini
  5. 05 · Fallback Gemini
2
model providers, with automatic fallback
3
memory layers: summary, database, vector
Read the case study

04 Data pipelines · LLM extraction

MCM: My Crypto Monitor

Backtests whether following crypto influencers on YouTube and Telegram would have made money or lost it.

Live platformJMD Tech SolutionsCo-developer of the core backend and AI pipeline (team of two)

  1. 01 · Discover Playwright
  2. 02 · Collect Videos · transcripts · messages
  3. 03 · Store MongoDB
  4. 04 · Extract DeepSeek
  5. 05 · Price Binance · CoinGecko
  6. 06 · Score ROI engine
40,000+
YouTube videos and transcripts collected
100,000+
Telegram messages processed
Read the case study

02 · How I work

Where one discipline runs out, the interesting work starts

My work usually starts where one discipline runs out. A chatbot needs memory, retrieval and a fallback for when the API goes down. A pressure mat needs firmware, filtering and a way to stop the language model from making things up. I like owning that whole chain and understanding every link in it.

Outside work I build and fly FPV drones, from 75 mm whoops to 7-inch long-range quads, 3D print parts in Fusion 360, and run my own Linux server. It is serving this page.

  1. Test, then choose

    I compared DeepSeek, Gemini and GPT on quality and cost before picking one for MCM, tested pipeline variants before splitting OCR from structuring for invoices, and tried resistor values until the sensor dividers fit.

    See: MCM: My Crypto Monitor, Purchase Invoice Extraction, FSP: Foot Pressure Mat

  2. Keep the model on a leash

    In FSP, deterministic rules decide the findings and the model only writes the explanation. In invoice extraction, anything that doesn't add up goes to a person, not into the spreadsheet.

    See: FSP: Foot Pressure Mat, Purchase Invoice Extraction

  3. Own it end to end

    From a flexible sensor matrix and an ESP32 to a 3D-printed enclosure, a backend and the report a user takes home.

    See: FSP: Foot Pressure Mat

03 · The lab

Things I build because I want to know how they work

Robotics · Flight control · 2024

Autonomous UAV

A custom 5-inch FPV drone made autonomous with a Raspberry Pi 5, ArduPilot and MAVLink.

  1. 01 · Mission DroneKit-Python
  2. 02 · Compute Raspberry Pi 5
  3. 03 · Link MAVLink over UART
  4. 04 · Control SpeedyBee F405 V3
  5. 05 · Fly 5-inch quad + GPS

Flown autonomously

Read more

Multi-agent AI · Real-time systems · 2026

AI Council

A brainstorming room where GPT, Gemini and Claude take turns discussing a problem with you, and remember past discussions.

  1. 01 · Client React + SSE
  2. 02 · API FastAPI
  3. 03 · Run Discussion runner
  4. 04 · Context Transcript + memory
  5. 05 · Recall ChromaDB
  6. 06 · Speak OpenAI · Gemini · Anthropic

In development

Read more
  • Offline RAG for private documents

    Fully offline document Q&A with a web UI, built on LocalGPT and Mistral 7B for a manager's private files. CPU inference was too slow, so I moved it to a GTX 1650 Ti. My first RAG system, where I learned chunking and retrieval.

    LocalGPT · Mistral 7B · RAG

  • Social-media automation

    Gathers trending posts and hashtags, has ChatGPT generate images and captions from that context, and posts and schedules to LinkedIn, Instagram and Facebook through their APIs.

    Playwright · OpenAI · Social APIs

  • Licence plate recognition

    Contour-based, CPU-only ALPR for gate entry that logs vehicle number and entry or exit time. Works best with a decent camera, good lighting and the right angle.

    OpenCV · Python

  • Devanagari character classifier

    A CNN trained on labelled Hindi Devanagari characters, reaching 90% accuracy; also explored AlexNet and GoogLeNet. B.Sc., 2021.

    CNN · Keras

  • This server

    A home Linux server running Nginx, Docker and Let's Encrypt. It hosts this site and AI Council, and will grow into my own self-hosted services.

    Ubuntu · Nginx · Docker

04 · Experience

3+ years building AI products

  1. Dec 2025 – Present

    AI Engineer · Innovizia Technologies

    • Own the core AI pipeline of My Best Self, a generative AI chatbot platform in beta on Google Play.
    • Independently built and deployed an invoice extraction platform now in use by a private firm.
    • Led end-to-end technical development of FSP, a foot-pressure screening product, from sensors to PDF report.

    Projects: My Best Self (MBS), Purchase Invoice Extraction, FSP: Foot Pressure Mat

  2. Sep 2023 – Dec 2025

    AI & ML Developer · JMD Tech Solutions

    • Co-developed the backend and AI pipeline of My Crypto Monitor in a two-person team, working directly with the client.
    • Collected and processed 40,000+ YouTube videos and 100,000+ Telegram messages; built the ROI engine.

    Projects: MCM: My Crypto Monitor

05 · Toolkit

Tools, and where I've actually used them

Applied AI & LLMs

  • OpenAI API
  • Gemini
  • DeepSeek
  • Qwen
  • OpenRouter
  • Prompt engineering
  • LLM evaluation
  • RAG

Used in: My Best Self (MBS), Purchase Invoice Extraction, MCM: My Crypto Monitor, FSP: Foot Pressure Mat

Retrieval & local inference

  • ChromaDB
  • Nomic embeddings
  • LangChain
  • vLLM
  • Llama models

Used in: My Best Self (MBS), AI Council

Backend & data

  • Python
  • FastAPI
  • REST APIs
  • SQL
  • PostgreSQL
  • MySQL
  • MongoDB
  • Redis
  • Celery
  • RabbitMQ
  • Alembic
  • Playwright

Used in: My Best Self (MBS), MCM: My Crypto Monitor, AI Council

ML & computer vision

  • PyTorch
  • TensorFlow
  • Keras
  • OpenCV
  • CNNs
  • OCR

Used in: Purchase Invoice Extraction, FSP: Foot Pressure Mat

Deployment

  • Docker
  • Linux / Ubuntu
  • Nginx
  • PM2
  • Git / GitHub

Used in: My Best Self (MBS), Purchase Invoice Extraction, AI Council

Embedded & hardware

  • C++ (ESP32 firmware)
  • ESP32
  • Arduino
  • Raspberry Pi
  • Multiplexed sensor arrays
  • Sensor interfacing
  • PCB assembly coordination
  • Soldering

Used in: FSP: Foot Pressure Mat, Autonomous UAV

Making & flight

  • Fusion 360
  • FDM 3D printing
  • ArduPilot
  • MAVLink
  • DroneKit-Python

Used in: FSP: Foot Pressure Mat, Autonomous UAV

06 · Background

Education, recognition and leadership

Education

  • MSc, Computer ScienceMithibai College, Mumbai · 2024CGPA 9.04 / 10
  • BSc, Computer ScienceM.V.L.U. College, Mumbai · 2022CGPA 9.65 / 10

Languages

English, Hindi, Marathi, Gujarati

Achievements

  • Research paper, IDEA 2K24Led the research and development of, and presented, “Scenery Drawings: A Pictorial Journey into Human Minds with CNN Insights on the Perceived Past, Present, and Future”, under the mentorship of Prof. Amol Joglekar, Mithibai College.
  • 1st Rank, Python CompetitionComputer Education Center (CEC), M.V.L.U. College, February 2022.
  • Data Science CertificationSix-semester programme at CEC, M.V.L.U. College covering machine learning models and CNNs. September 2022.

Leadership

  • Led the Training and Placement initiative at Mithibai College.
  • Student Council President for M.V.L.U. Vibes Fest 2022.
  • Head of Department (Graphics), Mithibai TechSpark.

07 · Contact

Building something that needs AI, a backend and maybe a soldering iron?

I'm happy to talk about roles, projects and collaborations.

jayantirathod56@gmail.com

GitHub