Heet Mehta

• Designed and built a real time benchmarking system in Go to evaluate scalable backend infrastructure for a production robot fleet, comparing Prometheus, VictoriaMetrics, Thanos, and Mimir under loads exceeding 782k time series.
• Engineered a containerized load generator with Bazel, deployed on Kubernetes with Helm, producing reproducible results across all configurations.
• Delivered a structured technical report whose findings directly informed an infrastructure migration decision.

• Contributing to open source secure systems research and tooling.

• Applied numerical computing in Python to fuse noisy sensor data from 500+ IoT devices using RSSI based trilateration and Kalman filtering, achieving 1 meter indoor positioning accuracy under real world interference.
• Built multithreaded real time data pipelines and automated simulation scripts, reducing evaluation time by 30% while maintaining correctness of statistical estimates.

• Built an ML powered diagnostics system in Python using Scikit-learn on AWS Lambda, applying statistical anomaly detection across sensor streams from 20+ installations and reducing critical failures by 40%.
• Designed scalable data ingestion pipelines integrating multi source sensor data via MQTT and AWS IoT Core, improving fault detection accuracy by 20%.

A self-hosted alternative to Celery/Sidekiq/Temporal: a Go scheduling engine with a DAG workflow orchestrator, real-time React dashboard, and Go + Python worker SDKs, deployable as a single binary or via Docker.

Aegis lets developers submit background jobs and multi-step pipelines (defined as a DAG in YAML/JSON) to a central server that distributes them to worker processes over a WebSocket protocol I designed. It handles the hard parts of a production job queue: jobs are leased to workers with a timeout so a crashed worker's job automatically gets reassigned; failed jobs retry with exponential backoff and jitter before landing in a dead-letter queue; workflow steps advance reactively (via an internal event bus) the instant their dependencies succeed, supporting fan-out/fan-in graphs rather than just linear chains. It runs as a single compiled binary with an embedded web dashboard, supports either embedded SQLite (zero config) or PostgreSQL (for multi-node deployments), and uses Postgres advisory locks for leader election so multiple server replicas can fail over automatically.

• Distributed job queue with priorities, retries with exponential backoff + jitter, and a dead-letter queue
• DAG-based workflow engine with cycle detection and automatic fan-out/fan-in advancement
• Real-time dashboard: live job table, DAG visualizer, throughput/latency charts, worker monitor
• Polyglot worker SDKs (Go and Python) speaking a shared WebSocket dispatch protocol
• Leader election (Postgres advisory locks) for multi-node high availability
• Single-binary deployment via Go's embed, plus Docker/Docker Compose for one-command setup

• Go
• TypeScript
• React
• PostgreSQL
• SQLite
• WebSocket
• Docker
• Python
• Tailwind CSS
• GitHub Actions

A comprehensive system for tracking, analyzing, and optimizing API usage: API key management with fine grained permissions, sliding window and token bucket rate limiting, real time metrics and error rates, per user analytics, full request logging, an interactive dashboard, and threshold alerts.

Three components: a FastAPI backend for auth, rate limiting, and logging; a Kafka consumer that aggregates metrics into ClickHouse; and a React dashboard for visualization.

• Python
• FastAPI
• Redis
• Kafka
• ClickHouse
• React

Mixpanel style product telemetry tool with a custom dashboard builder and funnel analysis views.

• React
• Zustand
• PostgreSQL
• ClickHouse

Auction platform on the Ethereum blockchain where users browse, bid, and transact on goods through encrypted exchanges backed by smart contracts.

• Solidity
• Web3.js
• Ethereum

Streamlit web app that analyzes uploaded retinal images with a pretrained ML model, a 25% accuracy improvement over baseline.

• Python
• Streamlit
• Scikit-learn

Deposit ETH for a receipt token, stake it for a second token, earn rewards, and vote on governance proposals: four Solidity contracts (deposit, staking, governance, and a stats/leaderboard contract) behind a Next.js frontend.

• Next.js
• TypeScript
• Solidity
• ethers.js

A minimal two-ledger UTXO simulator implementing hashed-timelock contracts and cross-chain atomic swaps, standard-library Python only. Includes a short paper covering the design and a correctness, liveness, and safety evaluation.

• Python

Trains Transformer and RNN language models at five sizes on ABC notation music converted from the Lakh MIDI Dataset, then fits a power law to validation loss vs. parameter count, the same question scaling laws research asks for text, asked of music instead.

• Python
• PyTorch
• Weights & Biases

Agentic AI engine with explainable multi variable scoring across 30+ inputs, improving decision accuracy by 25%. Designed for real time state sync via Socket.io supporting 50+ concurrent sessions. Awarded 3rd place at HackNYU 2026.

• Python
• Next.js
• TypeScript
• Node.js
• Snowflake

Applications of Machine Learning in Detecting Unethical Sources of Raw Materials in Supply Chains in the Cosmetic Industry.

DOI: 10.1109/ICSES63445.2024.10763049

Led a pro bono website redesign supporting workforce reintegration programs, improving navigation and accessibility for community impact.

Mentored 8 members across software and analytical projects and organized 5 technical workshops on algorithms and engineering fundamentals.