# Himansh Mudigonda | Founding ML & Backend Engineer

> [TL;DR] ASU MSCS grad (GPA 3.81). Founding Engineer at VelocitiPM. Experienced in Agentic AI (LangGraph/CrewAI), MLOps (SageMaker/MLflow), and Distributed Systems (Go/C++/Python/Kafka).

## Information & Contact
- **Name**: Himansh Mudigonda
- **Email**: himansh.m@velocitipm.com
- **Website**: [himudigonda.me](https://himudigonda.me)
- **GitHub**: [github.com/himudigonda](https://github.com/himudigonda)
- **LinkedIn**: [linkedin.com/in/himudigonda](https://linkedin.com/in/himudigonda)
- **Location**: Phoenix, Tempe, AZ, USA

## Site Navigation & Capabilities
This website is an interactive living portfolio powered by an AI-first architecture:
- **Ask HimmiAI**: A custom RAG-powered AI assistant available site-wide to answer questions about my background.
- **Interactive Command Bar**: Press `Cmd+K` (Mac) or `Ctrl+K` (Windows) to jump between sections.
- **Global Shortcuts**: `H` (Home), `T+K` (Toolkit), `E+X` (Experience), `P+R` (Projects).

## Core Values & Philosophy
I build systems anchored in truth and discipline. These values guide my engineering craft:
- **Hungry**: I chase problems, not job titles. From a Nature publication at 20 to shipping 21+ systems solo — I move fast because standing still feels like falling behind.
- **Humble**: I've been the founding engineer, the intern, and the student — in the same year. Every role taught me something the last one couldn't.
- **Sharp**: I think in systems, not tasks. I ask why the architecture needs something, who it serves, and what breaks when it scales.
- **Ownership**: When I take something on, it gets done. No handholding. I've shipped production systems solo that most teams would staff 4 engineers for.
- **Velocity**: Speed is a skill. I go from zero to production-grade fast — not by cutting corners, but by building the right thing first and iterating from a MVP.
- **Breadth with Depth**: I architect distributed systems in the morning and fine-tune LLMs in the afternoon. I go deep enough in each domain to make real decisions.

## Technical Toolkit

### Languages and Tools
- Python, Rust, JavaScript, TypeScript, SQL, Bash, Git, Docker, Kubernetes

### AI/ML
- PyTorch, TensorFlow, Keras, HuggingFace, Transformers, ONNX, OpenCV, scikit-learn, NumPy, Pandas, Computer Vision, Distributed Training, LangGraph, LiteLLM, Semantic Caching

### LLMs & Agentic Systems
- LangChain, LangGraph, CrewAI, A2A, RAG Architecture, Prompt Engineering, Fine-Tuning, Agentic AI, MCP, Inference Gateways, Multi-Provider Routing, MCP (Model Context Protocol)

### Backend & Orchestration
- FastAPI, Flask, gRPC, Kafka, Airflow, Pydantic, Celery, RabbitMQ, Redis, WebSockets, Node.js, Nest.js, Asynchronous Programming, GraphQL

### Cloud & IaC
- AWS, CDK, Lambda, SageMaker, Kinesis, S3, ElastiCache, SQS, SNS, Azure, Terraform, GCP, Vertex AI, BigQuery, Cloud Functions

### Databases & MLOps
- MySQL, PostgreSQL, DynamoDB, MongoDB, Redis, Elasticsearch, Pinecone, VectorDB, RDS, MLflow, CI/CD, Prometheus, Grafana, Distributed Tracing, OpenTelemetry, Performance Benchmarking

## Professional Experience

### AI Engineer @ Ron Turley Associates, Inc. (AKA RTA Fleet)
*2026-04-24 – Present | Glendale, AZ, USA*
- **1st AI hire** and 5th on the Innovation team — embedded in the special projects group driving company-wide AI transformation.
- Architected **FleetBeat**, RTA's proprietary Fleet Intelligence & Optimization Framework — delivered **$1M in hard cost savings in Q2 2026**, with a projected **$5M+ impact (+260%) by year-end**.
- Designed and owned **Fleet Success Dynamic UI**, a real-time personalized KPI surface giving fleet operators instant ad-hoc analytics — projected to reclaim **300,000+ analyst work hours** annually.

### Founding Machine Learning & Backend Engineer @ VelocitiPM LLC
*2025-06-01 – 2026-04-24 | Phoenix, AZ, USA*
- Architected a **30-agent AI orchestration engine** (LangGraph, CrewAI) on AWS AppRunner, automating 80% of PM workflows and increasing engineering throughput **4×**.
- Engineered an event-driven, async pipeline using **AWS Kinesis, SNS, and Lambda**, reducing P95 latency by **85%** and ensuring **99.98%** message delivery reliability.
- Built a production-grade **MLOps platform** with SageMaker Pipelines and MLflow, automating retraining gates and boosting release cadence **8×**.
- Implemented a robust **observability suite** (CloudWatch, X-Ray) that cut critical incident resolution time from 2 hours to **~10 minutes** via distributed tracing.
- Designed a **Redis-based memory layer** for agentic state persistence, slashing inter-agent latency by **60%** and supporting **2.5×** higher concurrent user throughput.

### Founding AI Engineer @ TimelyHero, Dimes Inc.
*2024-08-01 – 2025-06-01 | Remote - Tokyo, Japan*
- Spearheaded the migration of a legacy monolith to a **Java/Flask/gRPC microservice architecture** on AKS, scaling to support **100,000+** concurrent WebSocket sessions.
- Designed and deployed **Airflow-orchestrated RAG pipelines** (Pinecone, OpenAI, MongoDB), reducing data staleness from 48 hours to **30 minutes** and driving **$250K+** in new enterprise contracts.
- Standardized **Infrastructure-as-Code (IaC)** using Terraform and Kubernetes, reducing deployment downtime by **35%** and quadrupling the team's release velocity.
- Optimized high-throughput ingestion pipelines with **Kafka back-pressure handling**, stabilizing throughput at **5,000 msgs/sec** under peak loads of 1M+ events/hour.
- Implemented a comprehensive **monitoring stack** (Prometheus, Grafana) that improved proactive incident detection by **70%**, ensuring high availability.

### Machine Learning Intern @ Endimension Inc.
*2024-04-01 – 2024-08-01 | Remote - Tempe, AZ, USA*
- Trained and optimized large-scale vision models (TensorFlow, Keras) on **8TB of medical imaging data**, improving diagnostic mAP by **13%** and IoU by **17%**.
- Deployed **ONNX-quantized inference endpoints** on SageMaker, reducing P95 latency by **40%** with negligible AUC drift (<0.5%) for real-time diagnostics.
- Leveraged **SageMaker Spot Instances** and distributed training strategies to reduce GPU training costs by **25%** while accelerating iteration cycles by **1.8×**.
- Architected a **serverless inference pipeline** using AWS Kinesis and Lambda, ensuring **24/7 fault tolerance** and scalable processing for high-volume image streams.
- Configured a robust **model monitoring stack** (MLflow, Prometheus) to track drift and performance, reducing debugging time for production models by **60%**.

### Graduate Research Assistant @ JLiang Lab, Arizona State University
*2023-09-01 – 2024-05-01 | Tempe, AZ, USA*
- Developed state-of-the-art **multi-modal CV models** (Swin Transformer, DINOv2) on SageMaker GPU clusters, achieving a **24% increase** in rare-disease recall.
- Implemented **Fully Sharded Data Parallel (FSDP)** training on Amazon EKS (8× A100s), reducing model training time by **3.2×** and cutting cloud compute costs by **40%**.
- Engineered a **multi-node GPU job scheduler** using Slurm and EC2 Batch, automating provisioning and boosting cluster utilization by **35%** across shared workloads.
- Built distributed data pipelines with **AWS Glue and S3** to ingest and preprocess **8TB+** of multimodal medical datasets, improving data throughput by **2.6×**.
- Designed reproducible experiment tracking workflows with **SageMaker Experiments and MLflow**, reducing hyperparameter tuning time by **45%** and ensuring 100% reproducibility.

### Machine Learning Researcher @ SRM Advanced Electronics Laboratory
*2021-12-01 – 2023-07-31 | Amaravathi, AP, India*
- Developed a distributed **kernel regression pipeline** on Apache Spark (Java + MLlib), achieving a clinical-grade **MARD of 8.86%** for non-invasive glucose monitoring.
- Engineered a real-time IoT streaming system using **AWS Greengrass and IoT Core**, enabling ultra-low latency ingestion (**<200ms**) for 2,000+ daily sensor readings.
- Designed production-grade **ETL and data quality workflows** with Spark SQL, ensuring **95%+ signal integrity** across distributed edge devices.
- Optimized cloud-to-edge messaging protocols via **MQTT**, ensuring reliable data transmission and reducing packet loss by **15%** in unstable network environments.
- Co-authored and published this novel research in **Scientific Reports (Nature Portfolio)**, a Q1 journal, validating the system's clinical accuracy and architectural robustness.

## Selected Projects

### Protocol Battle Arena: High-Performance Benchmarking Suite (2026 Q1)
- **Stack**: Python 3.12, FastAPI, gRPC, GraphQL (Strawberry), WebSockets, SSE, Streamlit, Plotly, SQLAlchemy, psutil
- **Summary**: A sophisticated engineering lab for benchmarking modern web protocols. Features automated server orchestration, live resource tracking, and deep performance analytics for REST, gRPC, GraphQL, and streaming protocols.
- **Problem**: Comparing web protocols usually involves disparate tools and inconsistent environments. There is rarely a unified way to measure TTFB (Latency) and RPS (Throughput) while simultaneously monitoring system-level CPU and RAM pressure across different serialization formats.
- **Solution**: I built a robust **Control Tower** that manages five distinct protocol servers as background processes. I implemented a custom **Benchmarking Engine** that uses asynchronous `httpx` and `grpcio` to simulate high-concurrency "Siege" patterns. Results are visualized through interactive Plotly box plots that capture P50/P95 latency distributions.
- **Result**: The suite successfully identifies protocol breaking points under load, demonstrating gRPC's dominance in high-concurrency binary serialization versus REST's overhead. It provides a standardized environment for testing 1,000,000+ records with optimized SQLite batching.
- **Source & Demo**: [https://github.com/himudigonda/HandsOnWebProtocols](https://github.com/himudigonda/HandsOnWebProtocols)

### Zerobrew: Open Source Rust Systems Contribution (2026 Q1)
- **Stack**: Rust, Cargo, CLI, Parallel I/O, Git (Interactive Rebase)
- **Summary**: Implemented core **Multi-Formula Installation** for a high-performance Homebrew alternative in Rust. Optimized a sequential O(N) process into a parallelized batch operation.
- **Problem**: The tool originally only supported single-package installs. Users faced significant friction when setting up environments with multiple dependencies, as the CLI lacked a vector-based input handler for its three internal crates.
- **Solution**: I refactored `resolve_closure` to accept a slice of roots, allowing the dependency graph to resolve 100+ formulas simultaneously without redundant computations. I integrated the `?` operator for robust error bubbling and ensured the build passed a strict **-D warnings** Clippy check in CI.
- **Result**: Merged **PR #138**, closing a core bottleneck (Issue #18). The tool now supports batch installs, enabling it to compete with Homebrew’s standard UX while leveraging Rust’s safety and speed.
- **Source & Demo**: [https://github.com/lucasgelfond/zerobrew](https://github.com/lucasgelfond/zerobrew)

### HimmiRouter: Enterprise LLM Gateway & Workbench (2026 Q1)
- **Stack**: Python 3.12, FastAPI, LangGraph, LiteLLM, React, Shadcn UI, PostgreSQL (SQLModel), Redis, OpenTelemetry, MCP
- **Summary**: A production-ready **LLM Inference Gateway** and "Obsidian" themed AI Workbench. Features a state-machine router, atomic credit-based billing, and real-time distributed tracing for 80+ state-of-the-art models.
- **Problem**: Managing multiple LLM providers (OpenAI, Anthropic, Google) usually leads to fragmented API keys, lack of unified cost tracking, and no observability into routing logic. Standard proxies don’t handle complex stateful requirements like fallback routing or semantic caching at scale.
- **Solution**: I architected the core routing engine using **LangGraph** as a state machine to handle Auth -> Semantic Cache -> Cost-Aware Routing -> Provider Execution. I implemented **atomic row-level locking** in PostgreSQL to ensure credit integrity during streaming and integrated **OpenTelemetry/Jaeger** for full-trace visibility of the inference lifecycle.
- **Result**: The system unifies 84 models under a single OpenAI-compatible endpoint with **sub-millisecond routing overhead**. It features a "Shadow Mode" for real-time RLHF data collection and an **MCP Server** integration, allowing external agents like Claude Desktop to use the gateway as a tool.
- **Source & Demo**: [https://github.com/himudigonda/HimmiRouter](https://github.com/himudigonda/HimmiRouter)

### SuperSay: High-Performance Local AI Speech Engine (2026 Q1)
- **Stack**: Swift, AVFoundation, Python, FastAPI, ONNX Runtime, Kokoro-82M, Asyncio, macOS Engineering, Digital Signal Processing (DSP)
- **Summary**: A production-grade, **100% offline Text-to-Speech (TTS)** engine for macOS that eliminates the latency of local AI models through a parallelized inference-streaming architecture.
- **Problem**: Standard local TTS implementations are sequential, requiring the system to generate an entire audio file before playback begins. This results in multi-second "Time-to-First-Audio" (TTFA) delays that disrupt user flow. Furthermore, research papers are often cluttered with citations `[1, 2]` and page artifacts that ruin the narrative experience.
- **Solution**: I architected a **Priority-First Parallelization** engine using a hybrid Swift/Python IPC model. The system utilizes an **AsyncGenerator** that prioritizes the first sentence to minimize TTFA, while simultaneously firing background ONNX inference tasks for subsequent sentences. Audio is delivered via a **Chunked Binary Stream** over local HTTP, allowing for immediate hardware-buffer scheduling while the AI is still processing future text.
- **Result**: Benchmarked on Apple Silicon, the engine achieved an **instant-on TTFA of ~322ms** and a **Real-Time Factor (RTF) of 0.15**, generating audio 6.4x faster than real-time speech. The custom pipeline delivers a **1.6x parallel speedup** and successfully strips **99% of academic citations**, providing a seamless, distraction-free listening experience.
- **Source & Demo**: [https://github.com/himudigonda/SuperSay](https://github.com/himudigonda/SuperSay)

### Broski: Deterministic Task Runner for Polyglot Projects (2026 Q1)
- **Stack**: Rust, CLI, DAG / Graph Execution, Content-Addressed Caching, Live TUI, CI/CD, MIT
- **Summary**: A drop-in task runner written in **Rust** with Justfile-style simplicity, out-of-the-box caching, and an `--explain` flag that tells you exactly why a build reran.
- **Problem**: Make/Just-style runners rerun work blindly and offer no insight into *why* a task fired. Worse, a task that fails midway can leave the workspace in a poisoned, half-written state — turning rerun debugging into real wasted time for solo developers and small teams.
- **Solution**: Tasks declare their inputs/outputs (`@in` / `@out`); Broski hashes inputs to a content-addressed cache so unchanged work is skipped, and `--explain` surfaces the exact file that triggered a rerun (e.g. *"cache miss: input changed: tests/test_api.py"*). **Transactional output promotion** prevents poisoned workspaces on failure, and both **interactive** (`@mode interactive`) and **graph** execution are first-class. A live TUI (`broski tui ci`) renders DAG state, a log tail, ETA estimates, and cooperative Ctrl-C cancellation.
- **Result**: Released at **v0.7.0** with a 10-second `curl | bash` install and green CI — production-usable for local and small-team workflows, with `broski history` summarizing recent runs straight from the artifact store. Documented at [himudigonda.me/broski_docs](https://himudigonda.me/broski_docs).
- **Source & Demo**: [https://github.com/himudigonda/Broski](https://github.com/himudigonda/Broski)

### SuperZen: High-Performance Digital Wellness Guardian for macOS (2026 Q1)
- **Stack**: Swift, SwiftUI, SwiftData, AppKit, CGShieldingWindowLevel, Core Graphics, macOS Engineering
- **Summary**: A premium, **100% local** native macOS utility that combats Computer Vision Syndrome and fatigue by enforcing the **20-20-20 rule** through a context-aware, non-intrusive nudge system.
- **Problem**: Long screen sessions cause eye strain, poor posture, and fatigue, yet generic break timers are intrusive and easy to dismiss — so people ignore them and lose the health benefit entirely.
- **Solution**: A two-layer warning system precedes every break: a frosted **cursor-following pill** with zero-lag hardware tracking, then a data-rich corner alert with quick actions. When a break begins, SuperZen raises a full-screen shield at `CGShieldingWindowLevel` (the macOS login-screen level) across **all connected monitors**, offering blurred-wallpaper, custom-image, or mesh-gradient modes. Periodic **wellness pulses** cue blinking, posture, hydration, and affirmations, governed by **focus schedules** and **quiet hours**.
- **Result**: Shipped through **v1.1.7** with a comprehensive analytics engine (including a Focus Quality Score) — all data stored locally via **SwiftData**, so insights never leave the device.
- **Source & Demo**: [https://github.com/himudigonda/SuperZen](https://github.com/himudigonda/SuperZen)

### ytdld: Mission Control for Music & Video Downloads (2026 Q1)
- **Stack**: Python 3.11+, Textual TUI, yt-dlp, SponsorBlock, asyncio, macOS / Linux, MIT
- **Summary**: A TUI-first **mission control dashboard** for downloading YouTube content as MP3, MP4, or FLAC — powered by yt-dlp but built for humans, not flags.
- **Problem**: yt-dlp is powerful but CLI-heavy. Grabbing a clean 320kbps MP3 from a playlist, auto-skipping sponsors, and watching 8 parallel downloads in flight requires arcane flag combinations — and gives zero visual feedback on what's actually happening.
- **Solution**: Three interactive Textual screens guide the user from URL input to format selection to a **live parallel-download dashboard** (up to 8 concurrent workers). Output formats: **MP3 at VBR ~320 kbps**, MP4 with configurable quality caps, or **FLAC lossless**. SponsorBlock auto-skips or labels sponsor/intro/outro segments. A **download archive** makes re-runs fully idempotent. Browser-cookie passthrough unlocks age-gated and members-only content. Config persists to `~/.config/ytdld/config.toml`; a headless flag interface supports scripting.
- **Result**: Ships with a `curl | bash` one-liner, UV tool install, shell completions, and desktop notifications. Handles the full download lifecycle — URL input, quality selection, parallel fetch, archive dedup, SponsorBlock — from a single terminal session.
- **Source & Demo**: [https://github.com/himudigonda/ytdld](https://github.com/himudigonda/ytdld)

### IngestIQ: Enterprise Multi-Tenant RAG Platform (2025 Q4)
- **Stack**: FastAPI, Apache Airflow (Kubernetes Executor), RabbitMQ, ChromaDB (Distributed), OpenAI/Azure Embeddings, MLOps, Distributed Systems, Docker, PostgreSQL, PyTesseract/OCR, Redis Caching
- **Summary**: A high-throughput, event-driven **Enterprise RAG Platform** architected on FastAPI and Airflow, delivering **SOC2-ready data isolation**, millisecond-latency ingestion, and real-time semantic search for multi-tenant SaaS applications.
- **Problem**: Processing high-velocity, unstructured enterprise data (PDFs, DOCX, Images) at scale is prone to bottlenecks. The core challenge was building a system that guarantees **strict tenant isolation**, handles **corrupt files** gracefully without stalling pipelines, and ensures **idempotency** to prevent duplicate vector costs, all while maintaining <2s end-to-end latency.
- **Solution**: The system leverages an **async producer-consumer pattern** with RabbitMQ to decouple ingestion from processing. Airflow DAGs utilize **Dynamic Task Mapping** to parallelize OCR (Tesseract) and chunking across a Kubernetes cluster. A custom **Redis-based locking mechanism** ensures idempotency based on file content hashes. The query layer enforces security via **mandatory metadata filtering** and uses **Hybrid Search (Dense + Sparse)** for superior retrieval accuracy.
- **Result**: The platform successfully scaled to process **50,000+ documents/hour** with a 99.9% success rate. It reduced data availability latency from minutes to **<2 seconds**, enabling real-time RAG features for enterprise clients. The architecture passed a third-party security audit for **tenant isolation integrity**.
- **Source & Demo**: [https://github.com/himudigonda/IngestIQ](https://github.com/himudigonda/IngestIQ)

### Agentum-Framework (2025 Q4)
- **Stack**: Python, Agentic AI, LangGraph, OpenTelemetry, Redis, State Machines, Multi-Modal, FastAPI, Distributed Systems
- **Summary**: An open-source, **distributed framework** for orchestrating production-ready, multi-modal AI agents with **self-healing capabilities** and complex, stateful workflows.
- **Problem**: Productionizing AI agents is complex due to non-deterministic behaviors. Developers struggle with **distributed state management**, lack of observability into agent "thought processes" (the black box problem), and the difficulty of orchestrating **long-running, fault-tolerant** multi-agent collaborations.
- **Solution**: The framework uses a **hierarchical state machine** approach to manage complex agent trajectories, ensuring resilience and recoverability. It is built on an event-driven architecture, allowing agents to communicate asynchronously via a message bus. For observability, it integrates with OpenTelemetry to provide **distributed traces** of tool usage and model invocations. It supports **multi-modal inputs** and provides a unified interface for tools like web search and code execution.
- **Result**: The project is in active development, aiming to reduce boilerplate code for agentic systems by **60%**. Early benchmarks show support for orchestrating workflows with **10+ specialized agents** with sub-millisecond state transitions.
- **Source & Demo**: [https://github.com/himudigonda/agentum-framework](https://github.com/himudigonda/agentum-framework)

### CollabWrite (2025 Q3)
- **Stack**: Yjs, WebSockets, FastAPI, LangChain, Redis, PostgreSQL, Next.js, CRDTs, GPT-4, Vector Search
- **Summary**: A real-time, **AI-augmented collaborative editor** featuring **conflict-free synchronization** and context-aware writing assistance.
- **Problem**: Traditional editors lack real-time sync reliability and intelligent assistance, leading to **version conflicts** and inefficient workflows. Existing AI tools often lack context of the entire document history.
- **Solution**: I engineered a robust backend using **FastAPI and Redis Pub/Sub** to manage WebSocket connections at scale. The core synchronization relies on **Yjs CRDTs** to ensure eventual consistency across distributed clients. A **LangChain-powered RAG pipeline** feeds relevant document context into GPT-4, providing "aware" suggestions.
- **Result**: The application successfully handled **50+ concurrent users** in stress tests with zero data loss. User trials indicated a **45% reduction** in time-to-draft, validating the effectiveness of the AI-augmented workflow.

### FraudDetectX (2025 Q3)
- **Stack**: PySpark, PostgresSQL/ML, EvaDB, Graph Neural Networks, Transformers, ONNX, Real-Time Inference, Feature Store
- **Summary**: A **real-time fraud detection engine** leveraging **In-Database Machine Learning** and Graph Neural Networks for high-throughput transaction scoring.
- **Problem**: Traditional fraud systems suffer from **data movement latency** (ETL), allowing fraudulent transactions to slip through. Detecting complex fraud rings requires analyzing **entity relationships**, which is slow in standard SQL databases.
- **Solution**: By embedding ML models directly into PostgreSQL via EvaDB, the system performs **zero-copy inference** on streaming data. **Velox** accelerates query execution, while ONNX Runtime serves the GNN models. A **Feature Store** manages real-time aggregations for instant scoring.
- **Result**: The in-database architecture reduced inference latency by **70%** compared to ETL approaches. The GNN model improved fraud detection recall by **18%** by successfully identifying previously hidden collusion networks.
- **Source & Demo**: [https://github.com/himudigonda/FraudDetectX](https://github.com/himudigonda/FraudDetectX)

### Doppelgangerify (2025 Q2)
- **Stack**: LoRA, FLUX.1, PyTorch, Diffusers, ONNX, Dreambooth, Generative AI
- **Summary**: A personalized **Generative AI pipeline** that fine-tunes state-of-the-art diffusion models (FLUX.1) using **Low-Rank Adaptation (LoRA)** to create hyper-realistic user avatars.
- **Problem**: Standard text-to-image models fail to capture **specific identity traits**. Full model fine-tuning is computationally prohibitive and slow for individual users.
- **Solution**: The pipeline uses **LoRA** to inject user identity into the FLUX.1 attention layers without modifying the base weights. **Prior preservation loss** prevents overfitting. The trained adapters are exported to **ONNX** for optimized inference on consumer GPUs.
- **Result**: The system generates highly personalized images with a **35% improvement in FID scores** for likeness. Training time was reduced by **90%** compared to full fine-tuning, enabling a scalable "avatar-as-a-service" model.

### Gemma-3 Reasoning Training with GRPO (2025 Q2)
- **Stack**: GRPO, Gemma-3, PyTorch, RLHF, CUDA, GSM8k, LLM Alignment
- **Summary**: Fine-tuned the Gemma-3 model using **Group Relative Policy Optimization (GRPO)** to significantly enhance its **mathematical reasoning and logic** capabilities.
- **Problem**: LLMs often hallucinate on multi-step reasoning tasks. Standard Supervised Fine-Tuning (SFT) is insufficient for teaching **robust logic**. Reinforcement Learning (RL) is effective but often unstable and computationally expensive.
- **Solution**: The project leverages GRPO, a stable policy optimization algorithm, to align Gemma-3 with logical reasoning rewards. The training pipeline uses **gradient checkpointing** and **mixed-precision training (FP16)** to maximize throughput on limited compute.
- **Result**: The fine-tuned model achieved a **15% accuracy improvement** on the GSM8k benchmark. The GRPO approach proved **3x more sample-efficient** than standard PPO, demonstrating a viable path for efficient LLM reasoning alignment.
- **Source & Demo**: [https://github.com/himudigonda/reasoning-gemma3](https://github.com/himudigonda/reasoning-gemma3)

### SonicSherlock (2025 Q1)
- **Stack**: DSP, Audio Fingerprinting, PostgreSQL, FastAPI, Locality Sensitive Hashing
- **Summary**: A high-performance **audio recognition engine** utilizing **robust acoustic fingerprinting** and efficient database indexing to identify songs from noisy snippets.
- **Problem**: Identifying audio in real-time requires matching a short, noisy sample against a massive database. Brute-force comparison is impossible; a **highly optimized indexing strategy** is required.
- **Solution**: The system computes the **Short-Time Fourier Transform (STFT)** to generate spectrograms. It extracts "constellations" of peak frequencies and hashes pairs of peaks to create time-invariant fingerprints. These are stored in PostgreSQL and queried using a custom **temporal alignment algorithm**.
- **Result**: The engine identifies tracks from a **10,000+ song database** in **<3 seconds** with **95% accuracy**, even with significant background noise, demonstrating commercial-grade recognition performance.
- **Source & Demo**: [https://github.com/himudigonda/SonicSherlock](https://github.com/himudigonda/SonicSherlock)

### Beast Watch (2025 Q1)
- **Stack**: Computer Vision, Edge AI, YOLOv8, Gemini Pro Vision, FastAPI
- **Summary**: A **real-time wildlife safety system** combining **Edge Object Detection (YOLO)** and **Multimodal LLMs** to identify dangerous animals and provide immediate survival protocols.
- **Problem**: Human-wildlife conflict is increasing. In an encounter, seconds matter. Users need a tool that can **instantly identify** a threat and provide **context-specific safety advice** without relying on slow cloud round-trips for detection.
- **Solution**: The system uses a lightweight **YOLOv8 model** for sub-second detection of animal presence. Upon detection, the image is analyzed by **Gemini Pro Vision** to determine the exact species and behavior. The LLM then generates **location-aware safety instructions** (e.g., "Back away slowly," "Make noise").
- **Result**: The system identifies **50+ dangerous species** with **92% accuracy**. It delivers critical safety guidance in **<5 seconds**, potentially saving lives in conflict zones.
- **Source & Demo**: [https://github.com/himudigonda/BeastWatch](https://github.com/himudigonda/BeastWatch)

### ScripBuddy: AI Stock Analysis & Trading Assistant (2024 Q4)
- **Stack**: Python, Streamlit, Ollama LLM, HuggingFace, NLP Sentiment Analysis, RSI / MACD / Bollinger Bands, Plotly
- **Summary**: A stock-analysis and trading-assistance platform that fuses **technical indicators, news sentiment, and an LLM chat interface** into a single decision-support dashboard.
- **Problem**: Retail investors juggle fragmented tools — one for price action, another for fundamentals, another for news — making it slow and error-prone to form a confident, well-rounded view of a stock before acting.
- **Solution**: ScripBuddy presents interactive **candlestick charts** alongside technical indicators (Moving Averages, **RSI**, **MACD**, **Bollinger Bands**) and core fundamentals (market cap, sector, P/E, dividend yield). An NLP layer aggregates recent news and classifies market sentiment (Positive / Neutral / Negative), while a rule-based engine turns the indicators into actionable **Buy/Sell signals**. A chat interface — backed by an **Ollama** LLM with a **HuggingFace** model as an alternative — answers free-form questions and explains the analysis.
- **Result**: Delivers real-time, multi-signal analysis — price, technicals, fundamentals, sentiment, and trade signals — through one conversational interface, helping investors reach informed decisions faster.
- **Source & Demo**: [https://github.com/himudigonda/ML-trading-assistance-platform](https://github.com/himudigonda/ML-trading-assistance-platform)

### Ensemble Uncertainty Quantification for LLMs (2024 Q4)
- **Stack**: Bayesian Deep Learning, Uncertainty Quantification, LoRA, Ensemble Learning, PyTorch
- **Summary**: A framework for **Bayesian Uncertainty Quantification** in LLMs using **Deep Ensembles** and LoRA to detect hallucinations and improve reliability.
- **Problem**: LLMs are often **confidently wrong**. In high-stakes domains (medical, legal), knowing **when to trust the model** is as important as the answer itself.
- **Solution**: The project trains an **ensemble of lightweight LoRA adapters** on a frozen base model. By measuring the **disagreement (variance)** across the ensemble's predictions, the system quantifies epistemic uncertainty. This provides a calibrated "confidence score" for every generation.
- **Result**: The uncertainty metrics showed a **strong correlation (0.85)** with hallucination rates. The system successfully flags **incorrect answers** with high precision, enabling safer LLM deployment.
- **Source & Demo**: [https://github.com/himudigonda/EnsembleUQLLM](https://github.com/himudigonda/EnsembleUQLLM)

### MastoGraph - Mastodon (2024 Q3)
- **Stack**: Graph Neural Networks, NLP, Social Network Analysis, Llama3, NetworkX
- **Summary**: A **social intelligence platform** for the Fediverse, combining **Graph Algorithms** and **LLM-based NLP** to map influence and detect toxicity at scale.
- **Problem**: Decentralized networks are opaque. Understanding **community structure** and **content dynamics** requires analyzing both the graph topology and the semantic content of posts.
- **Solution**: The tool crawls the Mastodon instance graph. It uses **PageRank and Louvain Modularity** to identify influencers and communities. Simultaneously, a **Llama3 model** classifies post sentiment and toxicity. These signals are combined to generate a comprehensive "network health" report.
- **Result**: The tool analyzed **10,000+ nodes**, successfully identifying key opinion leaders and **toxic sub-communities** with **85% accuracy**, providing valuable insights for community moderators.
- **Source & Demo**: [https://github.com/himudigonda/mastograph](https://github.com/himudigonda/mastograph)

### FoR Audio: Fake or Real Speech Detection (2024 Q3)
- **Stack**: Audio Forensics, Deep Learning, RawNet, PyTorch, Self-Supervised Learning
- **Summary**: A **forensic audio analysis system** utilizing **RawNet and Self-Supervised Learning** to detect deepfake speech with high precision.
- **Problem**: Synthetic speech is becoming indistinguishable from human speech. Detecting deepfakes requires analyzing **micro-acoustic artifacts** that are invisible to the human ear.
- **Solution**: The system bypasses standard spectrograms, processing **raw audio waveforms** to capture phase information lost in conversion. It uses a **RawNet-based CNN** to extract features and a classifier to determine authenticity. Data augmentation (compression, noise) ensures robustness.
- **Result**: The model achieved **96% accuracy** on the ASVspoof benchmark, outperforming standard spectral baselines and demonstrating robustness against common audio post-processing.
- **Source & Demo**: [https://github.com/himudigonda/FoR_Audio](https://github.com/himudigonda/FoR_Audio)

### OpenForensics-DeepFake (2024 Q2)
- **Stack**: Computer Vision, Video Forensics, Swin Transformer, PyTorch, Temporal Analysis
- **Summary**: A **state-of-the-art deepfake video detector** leveraging **Swin Transformers** and temporal consistency analysis, achieving **98.59% accuracy**.
- **Problem**: Deepfake videos often exhibit **temporal inconsistencies** (flickering, jitter) that single-frame detectors miss. A robust system must analyze both spatial and temporal features.
- **Solution**: The system treats video as a sequence of patches. It uses a **Swin Transformer** backbone to extract hierarchical features. A **temporal attention module** aggregates information across frames to detect inconsistencies over time. Heavy **data augmentation** prevents overfitting to specific artifacts.
- **Result**: The model achieved **top-tier performance (98.59%)** on the OpenForensics challenge, validating the effectiveness of transformer-based architectures for video forensics.
- **Source & Demo**: [https://github.com/himudigonda/OpenForensics-DeepFake-Challenge](https://github.com/himudigonda/OpenForensics-DeepFake-Challenge)

### Classification & Localization Benchmarker (Pre 2023)
- **Stack**: MLOps, Computer Vision, PyTorch, Automated Benchmarking, Optuna
- **Summary**: An **extensible MLOps framework** for automated benchmarking and **hyperparameter optimization** of vision models across diverse datasets.
- **Problem**: Comparing model architectures is tedious. Researchers need a **standardized harness** to train, evaluate, and compare models like ResNet, ViT, and EfficientNet under identical conditions.
- **Solution**: The framework wraps popular libraries (timm, mmDetection) in a unified API. It automates the entire lifecycle: data loading, training loop, metric calculation (mAP, IoU), and result logging. It supports **distributed evaluation** for large-scale benchmarks.
- **Result**: The tool reduced experiment setup time by **50%**, enabling rapid iteration and fair comparison of model architectures for my thesis research.
- **Source & Demo**: [https://github.com/jlianglab/Thesis_mudigonda](https://github.com/jlianglab/Thesis_mudigonda)

### Otsu-Thresholding (Pre 2023)
- **Stack**: Computer Vision, Image Segmentation, Algorithm Design, Python
- **Summary**: A highly optimized, **vectorized implementation** of Otsu's Thresholding algorithm for automatic, unsupervised image segmentation.
- **Problem**: Understanding the core mathematics of computer vision algorithms is crucial for research. Standard library calls hide the implementation details.
- **Solution**: The project provides a clean, educational implementation of Otsu's method. It computes the **global threshold** that minimizes intra-class variance, effectively segmenting foreground from background without manual tuning.
- **Result**: The implementation is **mathematically verified** against standard libraries and serves as a clear, performant reference for understanding image segmentation fundamentals.
- **Source & Demo**: [https://github.com/himudigonda/Otsu-Thresholding](https://github.com/himudigonda/Otsu-Thresholding)

### NeuroLearn (Pre 2023)
- **Stack**: Neuro-AI, BCI, EEG Processing, Multi-Modal Learning, PyTorch
- **Summary**: A pioneering **Neuro-AI research initiative** decoding cognitive states from EEG signals using **Multi-Modal Deep Learning** to personalize education.
- **Problem**: Personalized education relies on understanding a learner's cognitive state. Traditional methods (quizzes) are reactive. **Direct neural decoding** offers a proactive, real-time measure of comprehension.
- **Solution**: The system records EEG data while subjects read text. A **1D-CNN** processes the EEG signals, while a **Transformer** encodes the text. These features are fused to predict the subject's "comprehension level." This feedback loop can dynamically adjust the difficulty of the material.
- **Result**: The model achieved **75% accuracy** in classifying cognitive load, providing a proof-of-concept for **Brain-Computer Interface (BCI)** driven adaptive learning systems.
- **Source & Demo**: [https://github.com/himudigonda](https://github.com/himudigonda)

### PopOS! Shell & Android AOSP ROM Development (Pre 2023)
- **Stack**: OS Development, Linux Kernel, Android AOSP, C, GNOME Shell
- **Summary**: Contributions to **core operating system components**, including the PopOS! window manager and custom **Android Kernel compilations** for extended device support.
- **Problem**: Stock operating systems often lack power-user features or drop support for older hardware. Community-driven development fills this gap.
- **Result**: My code improved the **workflow efficiency** of PopOS! users. My custom AOSP ROMs extended the life of the Redmi Note 7, serving a community of **hundreds of active users**.
- **Source & Demo**: [https://github.com/himudigonda/shell](https://github.com/himudigonda/shell)

### sCrAPTCHA & Archcraft Linux Contributions (Pre 2023)
- **Stack**: Cybersecurity, Python, Linux Customization, Shell Scripting
- **Summary**: Foundational work in **Web Security** and **Linux System Customization**, creating a custom CAPTCHA generator and enhancing the Archcraft distribution.
- **Problem**: Preventing bot attacks requires robust CAPTCHA systems. Linux distributions rely on community scripts for polish and usability.
- **Result**: The project demonstrated effective **security engineering principles**. My contributions to Archcraft helped streamline the user experience for a popular lightweight Linux distro.
- **Source & Demo**: [https://github.com/himudigonda/sCrAPTCHA](https://github.com/himudigonda/sCrAPTCHA)

## Publications & Research
- **Scientific Reports, Nature Publishing Group**: "Augmenting authenticity for non-invasive in vivo detection of random blood glucose with photoacoustic spectroscopy using Kernel-based ridge regression" (2024-04-09). Authors: P. N. S. B. S. V. Prasad V, Ali Hussain Syed, Mudigonda Himansh, Biswabandhu Jana, Pranab Mandal & Pradyut Kumar Sanki.
- **ICCCNT, IIT-Kharagpur, West Bengal, India. IEEE Xplore**: "Advancing Face Recognition Technology: A Comprehensive Analysis of Recent Breakthroughs and Emerging Research Frontiers" (2022-12-26). Authors: Medha Jha; Ananya Tiwari; Mudigonda Himansh; V. M. Manikandan.
- **ICITIIT, IIIT Kottayam, Kerala, India. IEEE Xplore**: "Quantifying the Impact: A Statistical Analysis of Open-Source Software Adoption and Its Critical Role in Modern Technology Ecosystems" (2022-04-01). Authors: M. Himansh and V. M. Manikandan.

## Certifications
### AI & Machine Learning Foundations
- **Supervised Machine Learning: Regression and Classification** - Stanford University (2024-08-28)
- **Neural Networks and Deep Learning** - DeepLearning.AI (2024-08-18)
- **Improving Deep Neural Networks: Hyperparameter Tuning, and Optimization** - DeepLearning.AI (2024-08-20)
- **Structuring Machine Learning Projects** - DeepLearning.AI (2024-09-05)
### Industry Specializations & MLOps
- **Google AI Essentials** - Google (2024-08-17)
- **Generative AI for Everyone** - DeepLearning.AI (2024-10-26)
- **MLOps Essentials: Model Development and Integration** - LinkedIn Learning (2025-02-06)
- **MLOps Essentials: Monitoring Model Drift and Bias** - LinkedIn Learning (2025-02-05)
### Professional Development
- **AI Model Development** - ASU School of Arts, Media and Engineering (2025-02-15)
- **Cross Functional Collaboration** - StarWeaver Group (2025-02-03)
- **Accelerate Your Learning with ChatGPT** - Deep Teaching Solutions (2024-10-13)
- **Data or Specimens Only Research** - Massachusetts Institute of Technology Affiliates (2023-09-01)
### Foundational Skills & Legacy
- **AI Foundations: Machine Learning** - LinkedIn Learning (2024-08-01)
- **Getting Started with Enterprise - grade AI** - IBM (2021-07-01)
- **Getting Started with Cloud for the Enterprise** - IBM (2021-07-01)

## Awards & Scholarships
- **Gold Medalist: Research Day** - School of Engineering, Arts and Sciences (2023-04)
- **Herbold ASU Graduate Scholarship** - Herbold Foundation: Bob Herbold, Ex-COO of Microsoft (2024-08-19 to 2025-08-19)
- **ASU Engineering Graduate Fellowship** - Ira A. Fulton Schools of Engineering (2023-07-21 to 2024-07-21)
- **SRM Merit Scholarship** - School of Engineering, Arts and Sciences (2019-06-01 to 2023-06-01)

## Education
- **Arizona State University**: Master of Science in Computer Science (AI/ML). **GPA: 3.81/4.0**. Activities: SoDA: Software Developers Association, ACM Student Chapter, Linux Users Group, The AI Society at ASU, Hindu YUVA. Coursework: Digital Image Processing, Image Analytics and Informatics, Statistical Machine Learning, Operationalizing Deep Learning, Cloud Computing, Data-Intensive Distributed Systems for Machine Learning, Advanced Operating Systems, Social Media Mining.
- **SRM University**: Bachelor of Technology in Computer Science & Engineering. Activities: Founder @ Inventors Village, Founder @ Research Clan, Board Member @SRM Student Council, Board Member @ SRM Entrepreneurship Cell, Member @ GDSC. Coursework: DSA in C, OOPs in Java, Database Management Systems, Computer Organization and Architecture, Introduction to Quantum Computations, Operating Systems, Data Warehousing and Data Mining, Computer Networks, Data Science, Software Engineering, Managing Innovation and Startups, Cloud Computing, Big Data Analytics, Machine Learning.

## Writing & Blog
- **BroSki: A Task Runner That Tells You Why It Reran** (2026-06-30): Why a Rust task runner needed transactional output promotion, content-addressed fingerprinting, and platform-aware sandboxing just to answer one question honestly: why did this rebuild? — [/blog/broski-deep-dive](https://himudigonda.me/blog/broski-deep-dive) — Tags: engineering, deep-dive, rust, developer-tools, broski
- **SuperSay: On-Device Text-to-Speech That Streams in Under 200ms** (2026-06-30): How SuperSay gets sub-200ms time-to-first-audio from a model that runs strictly one inference at a time, why INT8 quantization was tried and rejected with data, and the real production bugs (silent permission loss, a UI stuck at 0:00) that shaped the hardening sprint. — [/blog/supersay-deep-dive](https://himudigonda.me/blog/supersay-deep-dive) — Tags: engineering, deep-dive, swift, python, tts, supersay
- **SuperZen: Engineering a Native macOS Wellness App That Doesn't Get in Your Way** (2026-06-30): A deep dive into SuperZen's architecture — a single 1Hz heartbeat driving the entire app, privacy-first local-only telemetry, and the real bugs (App Nap stalls, midnight test failures, clipped shadows) that shaped the design. — [/blog/superzen-deep-dive](https://himudigonda.me/blog/superzen-deep-dive) — Tags: engineering, deep-dive, swift, macos, superzen