Mithran Sadasivam.

AI · Machine learning · Civic tech

Hi, I'm Mithran Sadasivam,

An AI Engineer and co‑founder.

I build machine learning models and the software around them, from PyTorch generative models to civic-AI for public procurement at Munifi Technologies. B.S. in Business Analytics, minor in Computer Science, University of San Francisco.

Selected projects

Select a project to see screenshots and diagrams. The two live demos use keyboard controls, so they work best on a desktop.

Munifi Vendor Management System

Civic-AI

Equity-procurement infrastructure that helps government and institutional buyers route contracts to certified small and diverse businesses.

Problem
Public buyers struggle to find and fairly evaluate certified small and diverse vendors, and must stay within rules like Prop 209.
Built
Separate agency-staff and vendor portals, race-neutral vendor scoring, automated compliance documentation and fraud-detection audit-chain tracking. A live proof of concept analyzes vendor-dispatch equity for San Francisco's Sidewalk Inspection and Repair Program.
Stack
Next.js, TypeScript, Tailwind CSS, Vercel

NYC 311 Streetlight Dispatch Platform

Full-stack · Geospatial

Automates reporting and dispatch for faulty-streetlight complaints. Awarded a USF Entrepreneurship & Innovation Hub microgrant.

Problem
Repair requests need to reach vendors quickly, and city rules (NYC PPB Rule 3-08) call for routing work to certified M/WBE businesses.
Built
A pipeline on NYC Open Data (311) with geospatial matching that maps each issue to the nearest certified vendor.
Stack
Python, NYC Open Data

CycleGAN Monet-Style Image Generation

Machine learning

A full CycleGAN built for the Kaggle Monet competition, producing a competition-ready submission.

Built
The complete generative adversarial network in PyTorch, with automatic CUDA, MPS or CPU detection, trained on GPU.
Stack
Python, PyTorch

Kinship

Web app

A family tree application that models multi-generational lineages as interactive, node-based maps.

Built
A deployed full-stack app with Google sign-in, join-by-code trees and approval requests. I used Claude and Cursor for debugging, refactoring and automated unit tests.
Stack
React, TypeScript, Django REST, Railway

Legion and Baymax

Local AI

Assistants that run entirely on a personal computer, with no cloud APIs or accounts.

Legion
A voice assistant that transcribes, reasons and speaks back sentence by sentence.
Baymax
A private health companion with a rule-based emergency gate that runs before the model, plus retrieval over clinical sources.
Stack
Python, faster-whisper, Ollama, Piper, RAG

Voltron and Kratos

Agents · Security

Two more local-first tools.

Voltron
A road-trip co-pilot where five role-specific assistants coordinate to answer questions, track fuel and warn the driver, served as a phone web app.
Kratos
A defensive security sentinel that audits your own code and configs in the background. Password checks use k-anonymity, so only a short hash prefix leaves the machine.
Stack
Python, Ollama, PWA, pywebview

Jager

Simulation · Autonomy

A concept for a legged search-and-rescue robot, built as a playable 3D simulation that runs in the browser.

Problem
A rescue robot has to decide where to search and whom to help first, not just how to drive.
Built
A robot modeled in code that finds hidden survivors from a noisy signal, keeps a Bayesian belief map of where they could be, and triages by time left. Drive it yourself or hand over to an autopilot that plans routes with A*. In recorded test missions the autopilot rescued 4.25 of 5 survivors on average.
Stack
JavaScript, Three.js, A*, Bayesian inference
Try the demo ↗

Megalodon Reclaimer

Simulation · Ocean cleanup

A concept for a robotic shark that collects plastic from the ocean while blending in with wildlife, built as a 3D simulation.

Problem
Large plastic and fishing debris drifts through the ocean, and a collector has to find it without scaring the animals around it.
Built
An autonomous shark, plus a three-headed Hydra variant, that sweeps with sonar, collects debris, and returns to a surface dock to unload and recharge. A stealth mode caps its speed so fewer fish are startled, and the HUD counts how many it disturbs.
Stack
JavaScript, Three.js, steering behaviors, boids
Try the demo ↗

Experience

Education

University of San Francisco. B.S. in Business Analytics, minor in Computer Science, Aug 2022 to May 2026. Coursework: Machine Learning, Data Structures and Algorithms, Statistical Analysis, Probability and Statistics.

Skills

Languages
Python, SQL, TypeScript, JavaScript, Java, HTML/CSS
ML and AI
PyTorch, scikit-learn, supervised, unsupervised and reinforcement learning, generative AI and LLMs, prompt engineering, Bayesian inference, path planning (A*)
Data and tools
Pandas, NumPy, PostgreSQL, MySQL, Snowflake, Tableau, Excel, Git
Engineering
React, Next.js, Three.js, Tailwind CSS, Vercel, API integration, AI-assisted development (Claude, Cursor)

Contact

Hiring, or want to talk about a project? Send me a note.

mithran.sadasivam@gmail.com