MaharshiBarot

AI Solutions Developer & Full-Stack Engineer — building intelligent products from idea to production.

langPython
langTypeScript
frontendReact
frontendNext.js
backendNode.js
backendFastAPI
backend.NET 8
mobileReact Native
aiOpenAI · Claude
aiRAG · Agents
cloudAWS · GCP
domainAI Agents
mobileCampus Ride Pooling
visionFace Recognition
agentsDataHub Steward Squad
agentsOverturn
agentsSeller Shield
domainComputer Vision
domainAutomation
domainFull-Stack
domainMobile Apps

About me

I'm an AI Solutions Developer and Full-Stack Engineer who ships intelligent products end-to-end — from autonomous agents and LLM systems to full-stack apps, mobile experiences and computer vision. I turn ambiguous ideas into working, tested software, and I upstream fixes to the open-source tools the industry runs on. Let's build something intelligent together.

Expertise

01

AI Systems & Agents

LLM products and autonomous multi-agent systems — RAG, tool use, grounding and evaluation — built with OpenAI, Claude, Gemini and Bedrock, and shipped to production.

02

Full-Stack Engineering

End-to-end web apps and APIs across React, Next.js, Node.js, Django, FastAPI and .NET 8 — from the data model to a deployed, tested UI.

03

Mobile Apps

Cross-platform apps with React Native and Firebase — real-time data, payments, identity verification and maps, released end-to-end.

04

Automation & DevOps

Background pipelines and cloud automation — Cloud Functions, schedulers, Pub/Sub, Docker and CI/CD — that keep working after you close the tab.

05

Computer Vision

Recognition systems with OpenCV and embeddings, hardened with blink-based liveness detection and anti-spoofing.

Selected Work

Project

01

Mobile · Full-Stack

Campus Ride Pooling

Stack

  • React Native
  • Firebase Auth
  • Firestore
  • Cloud Functions
  • Node.js
  • Stripe Payments

Architecture

  • React Native app
  • Firebase Auth
  • Firestore (real-time)
  • Cloud Functions
  • Stripe Payments + Identity

What it does

A full-stack campus ride-sharing app with real-time ride creation and chat, secure payments, identity verification, and route-aware ride matching.

  • Real-time ride creation and live chat between riders
  • Firebase Auth, Firestore and Cloud Functions backbone
  • Stripe payments plus Stripe Identity verification
02

Computer Vision · Backend

Face Recognition Attendance

Stack

  • Python
  • FastAPI
  • OpenCV
  • JWT Auth
  • RBAC

Architecture

  • Camera client
  • FastAPI service
  • OpenCV pipeline
  • 128-d embeddings
  • JWT + RBAC

What it does

An attendance platform that recognizes faces using 128-dimensional embeddings, defeats photo spoofing with blink-based liveness detection, and serves role-based analytics.

  • 128-d face embeddings for recognition
  • Blink-based liveness detection against spoofing
  • JWT authentication with role-based access control
03

AI Agents · Data Governance

DataHub Steward Squad

Stack

  • Python
  • Claude API
  • Model Context Protocol (MCP)
  • Multi-Agent Systems
  • DataHub

Architecture

  • DataHub (via MCP)
  • Catalog + lineage read
  • Five specialist agents
  • Chief Steward (Claude)
  • Approval-gated writeback

What it does

A five-agent system that reads real DataHub metadata over the Model Context Protocol, finds governance risks — missing owners, unclassified PII, lineage blast radius, failing quality assertions — and writes approval-gated fixes back to a live catalog, verified by re-reading.

  • Five specialist agents: catalog scout, lineage investigator, quality sentinel, stewardship writer, release captain
  • Reads real DataHub metadata over MCP and writes governed fixes back — then re-reads to prove they landed
  • Claude-powered Chief Steward reasons over grounded findings into an executive brief
04

AI Agents · Autonomous

Overturn

Stack

  • TypeScript
  • Gemini
  • Google GenAI SDK
  • Cloud Run
  • Firestore
  • Cloud Scheduler + Pub/Sub

Architecture

  • Patient uploads denial packet
  • Orchestrator (Cloud Run)
  • Intake → Diagnosis → Drafting → Filing
  • Gemini (GenAI SDK)
  • Firestore Memory Bank

What it does

An autonomous agent that fights wrongful health-insurance denials — it reads the denial packet, diagnoses why the claim was rejected, drafts a clause-cited appeal, files it, and keeps following up in the background for weeks until it's resolved.

  • Multi-agent pipeline: intake → diagnosis → drafting → grounding self-check → filing
  • Runs asynchronously in the background on a 14-day follow-up cadence, escalating after three attempts
  • Grounding agent blocks fabricated clinical facts and triggers automatic redrafts
05

AI Agents · AWS

Seller Shield

Stack

  • Python
  • Strands Agents SDK
  • Amazon Bedrock
  • Bedrock Knowledge Base
  • FastAPI
  • React

Architecture

  • Seller account-health feed
  • Monitor agent
  • RiskEvent
  • Evidence agent (Bedrock KB RAG)
  • Drafting agent

What it does

A four-agent AI team that protects small Amazon sellers — it watches account health in real time, catches suspension and return-fraud risk before it lands, and auto-drafts the evidence-backed appeal or SAFE-T reimbursement claim, citing the exact policy clause.

  • Four specialist agents — Monitor, Evidence, Drafting, Escalation — composed as tools under one orchestrator (Agents-as-Tools)
  • Detects suspension risk, ODR/OTDR drift and single-buyer return-fraud spikes, then routes to an advisory or a drafted claim
  • Bedrock Knowledge Base RAG retrieves the correct Amazon policy clause so every appeal cites real policy, never hallucinated text

Open Source

Contributions to the tools the industry builds on.

Real fixes upstreamed to major open-source projects — memory-safety, concurrency, type-inference and platform bugs. 2 merged (incl. Tesla & ROS 2) and 10 open across 7 organizations.

  • Tesla
  • Open Robotics · ROS 2
  • Meta
  • Google DeepMind
  • Waymo
  • PickNik · MoveIt
  • Supabase
  1. teslamotors organization avatar

    Tesla

    Tesla Vehicle Command SDK

    teslamotors/vehicle-command

    Merged

    tesla-http-proxy: exit 1 and report startup failures on stderr

    PR #524merged by @sethterashimaOct 2026

    1. Problem

      tesla-http-proxy could exit 0 and print nothing when it failed to start, so supervisors keyed on exit status (Docker / systemd on-failure, shell &&) treated a failed launch as a clean one.

    2. What I changed

      Split startup into small error-returning functions and made main a thin boundary that writes "Error: <reason>" to stderr and exits 1 regardless of log level.

    3. Impact

      Merged into Tesla's official Vehicle Command SDK by a Tesla maintainer — failed starts are now observable to process supervisors.

    • Go
    • TLS
    • CLI
    • DevOps
  2. ros-navigation organization avatar

    Open Robotics · ROS 2

    Nav2 · ROS 2 Navigation

    ros-navigation/navigation2

    Merged

    Fix BtActionNode cancellation when halted before goal acknowledgement

    PR #6445Issue #6426merged by @SteveMacenskiSep 2026

    1. Problem

      In Nav2's BtActionNode, halting during the window after a goal is sent but before the action server returns the goal handle could leave the in-flight goal uncancelled — an action-lifecycle race.

    2. What I changed

      Handled the halt-before-acknowledgement case so the goal is cancelled correctly once the handle arrives, with a regression test in test_bt_action_node.cpp.

    3. Impact

      Merged into main by Nav2's lead maintainer, Steve Macenski, and closes issue #6426 — a reliability fix in the flagship ROS 2 navigation stack.

    • C++
    • ROS 2
    • Behavior Trees
  3. facebook organization avatar

    Meta

    Pyrefly · Python type checker

    facebook/pyrefly

    Open PR

    Preserve Unknown when narrowing missing attributes

    PR #5346Issue #5333

    1. Problem

      When a missing attribute appeared in a narrowing condition, Pyrefly cached `object` instead of the error-origin `Unknown` from ordinary lookup, so later indexing produced a cascade of false errors.

    2. What I changed

      Recover missing lookup members as `Unknown` — preserving valid union members, positive refinement and genuine type errors — and treat reads of an error-`Unknown` as propagation to avoid duplicate diagnostics.

    3. Impact

      Open PR proposing recovery semantics for Meta's Pyrefly type checker (fixes #5333).

    • Rust
    • Type inference
    • Python typing
  4. facebook organization avatar

    Meta

    ThreatExchange · PDQ

    facebook/ThreatExchange

    Open PR

    [pdq][cpp] Prevent signed integer overflow in image downscaling

    PR #2026Issue #1979

    1. Problem

      In PDQ's C++ downscaling, row-major offsets and box-filter accumulators are `int`; for images where rows×cols exceeds INT_MAX this is signed-overflow UB, reachable from the public pdqHash256FromFloatLuma entry point and causing out-of-bounds access.

    2. What I changed

      Widened the offset and accumulator arithmetic to prevent the overflow and resulting OOB access, with a dedicated downscaling test.

    3. Impact

      Open PR hardening Meta's PDQ perceptual-hash library (ThreatExchange) against integer-overflow UB (fixes #1979).

    • C++
    • PDQ
    • Image hashing
    • Security
  5. facebook organization avatar

    Meta

    Mariana Trench · Security analyzer

    facebook/mariana-trench

    Open PR

    Fix false negative: taint propagation through static fields

    PR #277Issue #91

    1. Problem

      Mariana Trench missed taint flowing through static fields: a static write and a later read didn't share a storage location, so tainted values never reached the sink — a false negative.

    2. What I changed

      Added a StaticFieldMemoryLocation, stored static writes there, and resolved references through different declaring types to the same static field so taint is retained across the write.

    3. Impact

      Open PR closing an intraprocedural taint-tracking gap in Meta's Mariana Trench security analyzer (fixes #91).

    • C++
    • Static analysis
    • Taint tracking
  6. google-deepmind organization avatar

    Google DeepMind

    MuJoCo · Physics engine

    google-deepmind/mujoco

    Open PR

    Fix Studio native viewer on macOS: run the viewer on the main thread

    PR #3697Issue #3579

    1. Problem

      On macOS, AppKit services UI objects only on the main thread, but MuJoCo Studio's native launcher ran the viewer on a daemon worker thread — so the window was never created and the viewer failed to start.

    2. What I changed

      Added platform-aware launching that runs the viewer on the main thread on macOS while leaving Linux and Windows behaviour unchanged.

    3. Impact

      Open PR fixing the macOS native Studio viewer in Google DeepMind's MuJoCo physics engine (fixes #3579).

    • Python
    • macOS
    • Threading
  7. waymo-research organization avatar

    Waymo

    Waymax · Driving simulator

    waymo-research/waymax

    Open PR

    Fix scenario ID batching in the WOSAC notebook

    PR #96Issue #69

    1. Problem

      The WOSAC notebook decoded each scenario ID into a variable-length tensor before dense batching, so differing ID lengths failed in TensorFlow while equal-length IDs were silently concatenated.

    2. What I changed

      Kept IDs as strings until NumPy conversion, decoded each ID separately outside JAX, and compiled only the numeric simulator conversion — preserving the Python-string interface.

    3. Impact

      Open PR fixing scenario-ID batching in Waymo's Waymax autonomous-driving simulator (fixes #69).

    • Python
    • JAX
    • TensorFlow
  8. waymo-research organization avatar

    Waymo

    Waymax · Driving simulator

    waymo-research/waymax

    Open PR

    Fix configured simulation actors in the dm_env factory

    PR #94

    1. Problem

      make_sdc_dm_environment built configured IDM actors but omitted their parameter entries, so PlanningAgentEnvironment rejected any nonempty actor configuration before reset.

    2. What I changed

      Passed one None slot per parameter-free actor, preserving the factory API and constructor check, with regression coverage for batching, reset/step and invalid controls.

    3. Impact

      Open PR restoring configured simulation actors in Waymo's Waymax (dm_env factory).

    • Python
    • JAX
    • Reinforcement Learning
  9. waymo-research organization avatar

    Waymo

    Waymax · Driving simulator

    waymo-research/waymax

    Open PR

    Clarify rollout actors receive full state; add observation-based test

    PR #95Issue #73

    1. Problem

      RolloutCarry.observation's docstring said it was the actor's model input, but rollout() hands the actor the full simulator state — never that field — confusing users (reported in #73).

    2. What I changed

      Documented the actual intended design and added a regression test; no runtime behaviour changes.

    3. Impact

      Open PR clarifying rollout semantics in Waymo's Waymax (resolves #73).

    • Python
    • JAX
    • Docs & Tests
  10. teslamotors organization avatar

    Tesla

    React Native Camera Kit

    teslamotors/react-native-camera-kit

    Open PR

    fix(android): center scan frame after enabling and layout changes

    PR #822

    1. Problem

      Re-enabling showFrame on an already laid-out Android camera produced a 53×53px corner fragment because the overlay's measured dimensions were still zero.

    2. What I changed

      Computed frameRect from actual layout bounds in onSizeChanged() and used them when resizing the custom frame, with four Robolectric regression tests.

    3. Impact

      Open PR fixing the Android scan-frame layout in Tesla's React Native Camera Kit.

    • Kotlin
    • Android
    • React Native
  11. moveit organization avatar

    PickNik · MoveIt

    MoveIt 2 · Motion planning

    moveit/moveit2

    Open PR

    Fix use-after-free in servo_keyboard_input demo by joining the spin thread

    PR #3850

    1. Problem

      The servo_keyboard_input demo could read freed memory because its ROS spin thread outlived the resources it referenced.

    2. What I changed

      Join the spin thread on shutdown so it stops accessing freed memory — a clean lifetime fix for the demo.

    3. Impact

      Open pull request under review in MoveIt 2, the standard ROS 2 manipulation framework.

    • C++
    • ROS 2
    • MoveIt Servo
  12. supabase organization avatar

    Supabase

    Supabase Studio

    supabase/supabase

    Open PR

    fix(studio): show removed edge functions in branch merge diff

    PR #44194Issue #44175

    1. Problem

      When an edge function was removed from config.toml on a preview branch, Studio's Edge Functions merge tab showed "No changes detected", hiding the deletion so the function looked still-deployed.

    2. What I changed

      Fixed useEdgeFunctionsDiff so removed or disabled functions surface as a deletion diff on the merge page, with a regression test.

    3. Impact

      Open PR fixing branch-merge edge-function diffs in Supabase Studio (fixes #44175).

    • TypeScript
    • React
    • Next.js

The Journey

Experience & education.

Experience

  • Aug 2026 – Present

    Teaching Assistant

    Texas A&M University–Corpus Christi · Corpus Christi, TX

    • Strengthened undergraduates' command of algorithmic problem-solving and efficient data organization across a core Data Structures course by leading labs, live code walkthroughs, and office-hour support.
    • Built first-year students' programming fundamentals across an Introduction to Problem Solving course by teaching computational-thinking logic and guiding them from pseudocode to clean, working code.
  • Jan 2026 – Jun 2026

    Software Developer

    Meeka Consultancy LLC · United States

    • Drove API development for end-to-end frontend and backend features across the full software development lifecycle by reproducing, triaging, and resolving production defects to closure.
    • Strengthened software quality across 12 collaborative software releases by conducting code reviews, targeted quality assurance testing, and DevOps configuration management.
  • Jun 2025 – Nov 2025

    AI & LLM Product Developer Intern

    Treta Gen. · India

    • Conducted applied research across ambiguous customer problems by applying LLMs, Python scripting, and machine learning experimentation to test hypotheses in Linux/Unix environments.
    • Accelerated production AI product launches across the full development lifecycle by applying rapid prototyping, ML evaluation, and root-cause debugging in DevOps deployment practices.

Education

  • Jan 2026 – Dec 2027Ongoing

    Master of Science, Computer Science

    Texas A&M University–Corpus Christi

    • GPA 3.5
  • Sep 2022 – Jun 2025

    Bachelor of Engineering, Computer Engineering

    Pandit Deendayal Energy University

    • GPA 3.2

Technical Ecosystem

The tools, and what they built.

Hover a technology to light up the projects it powers — or a project to see its stack.

Languages

  • C#
  • C/C++
  • Java
  • JavaScript
  • MATLAB
  • PHP
  • Python
  • SQL
  • TypeScript
  • Bash

Frameworks

  • Angular
  • ASP.NET
  • D3.js
  • Django
  • Express.js
  • FastAPI
  • Flask
  • HTML/CSS
  • Next.js
  • PyTorch
  • React.js
  • React Native
  • Selenium
  • Spring
  • Tailwind CSS

Cloud & AI Tools

  • AWS
  • GCP
  • Azure
  • Docker
  • Kubernetes
  • Terraform
  • Linux/Unix
  • Git

Databases

  • DynamoDB
  • MongoDB
  • MySQL
  • Microsoft SQL Server
  • NoSQL
  • Oracle
  • PostgreSQL
  • Power BI
  • Redis
  • Elasticsearch
  • Firestore

Let's Talk

Let's build something intelligent.

Open to roles and collaborations where AI meets real products. If you're hiring or have an idea worth shipping, my inbox is open.