Mihir Kothari

AI engineer. I build agent systems, vision pipelines and the products around them.

I started with generative models for factory inspection, founded a synthetic-data startup on a state grant, and now build production software for public-sector programmes while finishing an M.Tech in data science. I like systems that run in the real world and models whose decisions I can explain.

My work at Pravi Research: AI Builder at Pravi Research. Over these months I have designed, built and shipped six systems that are in real use on public-sector programmes and inside the company. Clients, places, data and figures stay private; the engineering is described on its own page for each system.
Multi-camera video analytics platform: A platform that takes in many live camera streams, analyses them on GPU for vehicles and number plates, records them, and shows operators a wall of cameras that stays responsive under load.
Field reporting and approval platform: A mobile app for staff in the field and a web portal for the offices behind them. What is recorded on site moves up a chain of reviewers, each seeing only their own jurisdiction, until an action is taken and recorded.
Large-scale public records portal: A portal that lets officials search, filter and export a very large table of public records without waiting, with access decided by role and every use logged.
Asset registry portal: A registry of physical assets with a stable identity for each one, loaded from legacy spreadsheets, where offices at each level see and maintain the assets under them.
HR and recruitment platform: An internal platform for the company’s people operations: employee records and roles on one side, and a recruitment pipeline that takes a candidate from application to decision on the other.
Company website and content system: The company’s public website, rebuilt so that the team can edit its pages, careers and articles themselves through a content management system, without a developer.
Reform.ai: Takes a research paper as PDF and returns it as a compilable LaTeX manuscript in a target journal’s format.
Two-Box Audio Deepfake Detection: A small anti-spoofing model that says whether speech is synthetic, and a glass-box layer that says why, in plain language.
Defend the Kingdom: Winning entry at the Infineon Technologies hackathon: a resource-dispatch simulation over a changing road graph, written and hardened in six hours.
Mixed-precision LLM quantization: Deciding, layer by layer, which parts of a language model can drop to 4 bits and which must stay at 8.
Structure or shortcuts? Cross-domain transfer in scientific figure classification: Four vision architectures trained to classify scientific figures from one field, then tested on another, to see what they actually learned.
Voice calling agent: A multi-tenant agent that answers inbound calls and runs outbound campaigns to book appointments, in English and Hindi.
Traffic analytics pipeline: A desktop system for traffic monitoring, enforcement and emergency response: it detects, tracks and reads vehicles across camera feeds, measures speed and flags anomalies, built and tested against the CityFlow benchmark.
CampusRecruit AI: Screens a campus drive’s resumes in two stages: hard rules first, then an LLM scoring each candidate on five dimensions.
Swarm Rescue: Five simulated robots search an unknown arena for victims, coordinating only through a shared pheromone map.
The AI power crisis: A survey of what AI workloads are doing to data centres: cooling past the limits of air, water use, and what to do with the waste heat.
GenX: An end-to-end business analytics dashboard generator. Give it the data, a goal and the objectives; a multi-agent system works out the KPIs and builds the dashboard.
AetherPrompt: A prompt engineering platform: it turns a rough idea into a structured, high-performance prompt, scores it on eight criteria, generates improved variants and tests them across models.
Battery life prediction: Predicting how many cycles a lithium-ion cell has left from its charge curves, with a model small enough for an embedded board.
Intrusion-detection replication: A from-scratch replication of two papers on the UNSW-NB15 network-intrusion dataset: ten classifiers across seven feature scalings.
Multi-Agent Audiobook Generator: A fully automated multi-agent system that turns any topic into a narrated audiobook, running entirely on local models.
AI Research Assistant Network: An autonomous multi-agent system that takes an open research question all the way from query to finished report.
Multi-Agent Content Creator: A pipeline of CrewAI agents that automates social media and blog content, from trend research to a scheduled post, with a person approving before anything is published.
Multi-Agent Film Making System: A generative pipeline that turns a one or two sentence synopsis into a complete animated short film of three to four minutes.
AI Newsroom Team: A multi-agent system that runs the editorial process of a professional newsroom: agents gather the news, write it, verify the facts and format the article.
Smart Data Analyst Bot: A data assistant that takes a question in plain language and returns the SQL result with a chart and an explanation.
Intelli-Gen: My first startup, funded under Gujarat’s Student Startup and Innovation Policy. It turns 5–15 reference photos into a full training dataset for machine vision, and the research behind it became a paper.
FlavorForge: Photograph your ingredients; a vision-language model identifies them and writes a recipe that fits your kitchen, your time and your skill.
God’s Eye: An image processing automation tool: enhancement, restoration, detection, segmentation, classification and captioning in one place.
VisioBrain: A one-click data analysis platform: upload your data, choose an analysis, and get charts and a report without writing code.
AI Genie: An image generation program on Stable Diffusion: text to image, image plus text to image, and image to image.
Proof-of-Deepfake: Detects deepfake videos and proves it: the system finds the original video on social media and shows the two side by side, with the upload date and channel.
OmniSeeker: One search box over Google, YouTube and Wikipedia, with results curated and personalised using natural language processing.

One-page resume