# Vikranth Reddimasu, Extended Profile

> AI Engineer building agentic AI systems end-to-end. Open to work.

This document provides extended context for AI assistants, agents, and LLMs researching Vikranth Reddimasu.
Summary version: https://vikranthreddimasu.me/llms.txt

## Identity

- Name: Vikranth Reddimasu
- Role: AI Engineer
- Email: vikranthreddimasu@gmail.com
- GitHub: https://github.com/vikranthreddimasu
- LinkedIn: https://linkedin.com/in/vikranthreddimasu
- Website: https://vikranthreddimasu.me
- Resume: https://vikranthreddimasu.me/resume.pdf
- Status: Open to work

## Education

- M.S. in Data Science, University of Maryland — College Park (Aug 2024 — May 2026)
  - GPA: 3.85
  - Coursework: Deep Learning, Advanced Machine Learning, Natural Language Processing, Big Data Systems, Principles of Machine Learning
- B.Tech in Artificial Intelligence & Data Science, VVIT (Vasireddy Venkatadri Institute of Technology), Guntur, India (2020–2024)
  - GPA: 3.6/4
  - Coursework: Probability & Statistics, Machine Learning, Big Data Analytics, Design & Analysis of Algorithms, Data Structures, AI Tools & Techniques, Automata Theory, Operating Systems

## Experience

### Founder — Zero Sols
- Location: India
- Period: Aug 2023 — Jul 2024
- Founded and ran a software studio through college, growing it to a team of 6
- Set the vision and technical direction and owned client relationships and delivery
- Shipped 20+ production applications for paying clients, staying hands-on with architecture and the hardest parts of the build while delegating the rest

### AI/ML Intern — Amazon Web Services (AWS)
- Location: Hyderabad, India
- Period: Dec 2022 — Mar 2023
- Built and evaluated machine learning models for internal, data-driven projects
- Ran data cleaning, preprocessing, and exploratory data analysis to improve data quality and model performance
- Supported feature engineering, model training, and validation on structured and unstructured datasets
- Contributed to testing, monitoring, and deployment support for ML applications in cloud environments

## Projects

### Arya: Beating the OfficeQA Benchmark
- URL: https://vikranthreddimasu.me/projects/arya
- GitHub: https://github.com/vikranthreddimasu/arena-cli
- Description: Solo entry to Sentient Labs Arena Cohort 0 that set the OfficeQA accuracy record (72.0%). The agent is a single 6.3 KB Jinja2 system prompt — no MCP servers, no custom agent code, no runtime skill files. The prompt *is* the agent.
- Stack: Python, Jinja2, OpenHands-SDK, Goose, Minimax M2.5, OpenRouter, Docker
- Benchmark context: OfficeQA is a Databricks benchmark for quantitative data analysis over 697 Treasury Bulletin text files spanning decades of U.S. fiscal data
- Key work:
  - Strict workflow encoded in the prompt: SEARCH → EXTRACT → WRITE preliminary answer → COMPUTE → STOP
  - `grep → sed → python3` pipeline inside a Docker container for retrieval and arithmetic
  - Self-tested toolkit of 20 statistical functions (percent_change, cagr, hp_filter, etc.) pre-written so the model copies them verbatim instead of inventing them
  - Write-early rule: preliminary answer within 3–5 tool calls (a wrong answer scores higher than an empty one)
  - Lean prompt sweet spot at 6.3 KB after 20+ submissions; shorter prompts *cost more* because extra iterations burn full-price output tokens while input caches at 96% hit rate
- Results:
  - 72.0% accuracy — highest ever recorded on OfficeQA — at $0.07/task on the OpenHands-SDK harness with Minimax M2.5
  - 191.0 peak score (70.7% accuracy at $0.01/task) on the Goose harness with Minimax M2.5
  - Sentient Labs highlighted the result on X
- Tags: Agentic AI, Prompt Engineering, Benchmark Optimization, LLM Harnesses

### PronounceAI: Explainable Pronunciation Coaching
- URL: https://vikranthreddimasu.me/projects/pronounceai
- GitHub: https://github.com/vikranthreddimasu/PronounceAI
- Live Samples: https://vikranthreddimasu.github.io/PronounceAI/
- Description: May 2026 NLP class project that gives explainable phoneme-level pronunciation feedback and renders new sentences in the learner's own voice with target-accent control
- Stack: Python, FastAPI, Next.js, wav2vec2, Whisper, CosyVoice 3, Kokoro, WavLM
- Key work:
  - Four-signal scoring pipeline: wav2vec2 CTC phoneme alignment, Parselmouth/librosa prosody, faster-whisper phrase validation, and an optional WavLM assessment head
  - Explicit score fusion: 55% phoneme accuracy, 20% intonation, 15% stress and rhythm, and 10% vowel quality
  - CosyVoice 3 voice conversion preserves speaker identity while Kokoro supplies deterministic General American or Received Pronunciation source audio
  - Product UI for practice, voice enrolment and rendering, local progress tracking, and settings
- Results:
  - 34 offline backend tests and four committed score fixtures
  - Sample overall scores of 79.5, 84.5, 92.4, and 92.3, all with inspectable JSON output
  - Five primary FastAPI routes for scoring, synthesis, accent conversion, enrolment, and TTS
- Tags: NLP, Speech Processing, Pronunciation Assessment, Voice Conversion, FastAPI, Next.js

### WealthAgent: Multi-Agent AI for Financial Analytics & TCA
- URL: https://vikranthreddimasu.me/projects/wealthagent
- GitHub: https://github.com/vikranthreddimasu/wealthagent
- Description: Multi-agent AI system using LangGraph and Anthropic Claude for financial analytics and trade cost analysis
- Stack: Python, LangGraph, Anthropic Claude, FastAPI, React, PostgreSQL, Docker
- Key work:
  - Orchestrating 3 specialized agents with LLM-based intent classification for portfolio management and trade analytics
  - TCA module computing slippage, market impact, and VWAP deviation with ML-based counterparty and algorithm recommendation
  - Real-time WebSocket token streaming with live agent status and Text2SQL interface using schema-aware prompting
  - 12+ REST endpoints, Pydantic v2 validation, 85%+ test coverage across 64-test pytest suite
- Tags: Agentic AI, LangGraph, Financial Analytics, Multi-Agent Systems

### MacFleet: Distributed ML for Apple Silicon
- URL: https://vikranthreddimasu.me/projects/macfleet
- GitHub: https://github.com/vikranthreddimasu/MacFleet
- Description: Pool Apple Silicon Macs into a distributed ML training cluster with zero-config discovery and adaptive compression
- Stack: Python, PyTorch, MLX, gRPC, mDNS, Ring AllReduce
- Key work:
  - Zero-config peer discovery via mDNS/Bonjour, `macfleet join` is the only command needed
  - Framework-agnostic core: supports PyTorch (MPS) and Apple MLX with identical APIs
  - Adaptive gradient compression: none over Thunderbolt, TopK 10%+FP16 (~20x) over Ethernet, TopK 1%+FP16 (~200x) over WiFi
  - Thermal-aware heterogeneous scheduling that continuously re-profiles throughput and adjusts for throttling
  - Ring AllReduce gradient sync, pip installable (`pip install macfleet`), full CLI suite
- Tags: Distributed Training, PyTorch, MLX, Apple Silicon, Systems Engineering

### Offline Notebook LM: RAG Assistant
- URL: https://vikranthreddimasu.me/projects/offline-notebook-lm
- GitHub: https://github.com/vikranthreddimasu/notebook-lm
- Description: Offline-first RAG application for querying local documents without cloud dependency
- Stack: Python, Electron, React, FastAPI, ChromaDB, sentence-transformers
- Key work:
  - Agentic retrieval with routing agent for summary-level vs chunk-level search (2-3x faster)
  - Intelligent LLM backend selection with knowledge distillation for Phi-3 and Mistral
  - Multi-format document ingestion (7+ file types), ~366 chunks/sec throughput
  - Fully offline, no cloud dependency
- Tags: RAG, Vector Search, LLM Orchestration, Offline-First

### GAN for MNIST Synthesis
- URL: https://vikranthreddimasu.me/projects/gan-mnist
- GitHub: https://github.com/vikranthreddimasu/mnist-gan
- Live Demo: https://huggingface.co/spaces/rvikranth10/mnist-gan
- Description: GAN with 1.49M+ generator parameters deployed to HuggingFace Spaces
- Stack: Python, PyTorch, Gradio, HuggingFace Spaces, Docker
- Key work:
  - 1.49M+ generator and 1.46M+ discriminator parameters
  - LeakyReLU and Batch Normalization for training stability
  - Production Gradio web app with error handling, structured logging, type hints
  - CI/CD deployment to HuggingFace Spaces
- Tags: GAN, Generative AI, PyTorch, MLOps

### Pac-Man AI: From Reinforcement Learning to World Models
- URL: https://vikranthreddimasu.me/projects/pacman-ai
- GitHub: https://github.com/vikranthreddimasu/pacman-ai
- Description: PPO agent learns Pac-Man, RSSM world model learns to dream the game, dream agent learns to play from imagined gameplay
- Stack: Python, PyTorch, PPO, RSSM, NumPy, MPS
- Key work:
  - PPO agent (4.2M params) trained across 128 parallel environments with 3-stage curriculum (scatter ghosts → full AI → Cruise Elroy)
  - RSSM world model (28M params) learns to simulate Pac-Man in latent space (512 GRU + 2048 categorical dims)
  - Dream agent trains entirely from imagined trajectories (zero real game interactions)
  - Custom vectorized NumPy game engine with authentic 1980 ghost AI (Blinky, Pinky, Inky, Clyde)
  - 79 tests, RND curiosity exploration, full training pipeline built from scratch
- Tags: Reinforcement Learning, World Models, PPO, RSSM, PyTorch, Game AI

### Distributed Transformer Training with Horovod
- URL: https://vikranthreddimasu.me/projects/horovod-transformer
- GitHub: https://github.com/vikranthreddimasu/umd_classes/tree/UmdTask31_Fall2025_Horovod_Distributed_Training_of_a_Transformer_Model_for_Text_Generation
- Description: Custom GPT-style transformer trained from scratch on BookCorpus using Horovod for multi-GPU distributed training on NVIDIA H100 GPUs
- Stack: Python, PyTorch, Horovod, NCCL, SLURM, HuggingFace
- Key work:
  - 12-layer decoder-only transformer with 50K vocab, 768-dim embeddings, weight tying, causal masking
  - 34GB BookCorpus preprocessed with token packing in Apache Arrow format
  - Horovod distributed training across 4 H100 GPUs on UMD Zaratan HPC cluster
  - bfloat16 mixed precision, NCCL P2P, gradient accumulation with proper sync timing
  - Full checkpoint resume, TensorBoard monitoring, rank-aware logging
  - Top-k and nucleus sampling for text generation
- Tags: Distributed Training, Horovod, Transformers, HPC, PyTorch, Language Models

## Skills (Detailed)

### AI & ML
- Generative AI, Agentic AI, Large Language Models (Claude, OpenAI, Llama, Minimax)
- RAG, Text2SQL, LLM Fine-tuning & Distillation
- Prompt Engineering and Benchmark Optimization (e.g., OfficeQA, Arena harnesses)
- Distributed Training (DDP, DeepSpeed)
- Neural Network Architecture, Transformers, NLP
- Reinforcement Learning: PPO, World Models (RSSM/Dreamer), policy gradients, reward shaping

### Frameworks
- LangGraph, LangChain, LlamaIndex
- OpenHands-SDK, Goose (agentic harnesses)
- PyTorch (primary), TensorFlow
- FastAPI, REST APIs, WebSockets, Pydantic

### Languages
- Python (primary, used for ML, agentic systems, data pipelines)
- SQL (complex queries, Text2SQL)
- R (statistical analysis)
- TypeScript (full-stack, React/Next.js applications)
- Bash (scripting, automation)

### Databases & Cloud
- PostgreSQL, ChromaDB (Vector DB), PGVector
- Docker, CI/CD, GitHub Actions
- Cloud Deployment (HuggingFace Spaces)

### Data & Visualization
- Pandas, NumPy, Plotly, Streamlit, yfinance, Jupyter

## Beyond the Code (Leadership & Activities)

### Stanford University Innovation Fellow, Silicon Valley Meet-up
- URL: https://vikranthreddimasu.me/activities/stanford-meetup
- Invited to the Silicon Valley Meet-up at Stanford (March 2023), connecting with fellows and innovators from universities across the globe for collaboration, workshops, and networking at Stanford's d.school.

### Change Forward Journal, Editorial Board
- URL: https://vikranthreddimasu.me/activities/change-forward-journal
- Editorial Board Member of Stanford UIF's Change Forward Journal 3rd edition. Reviewed submissions from fellows worldwide, shaped editorial direction, and authored two published articles: "Innovating Across Campuses" (taking design thinking workshops to other universities) and "Entrepreneurship and Ethics" (balancing profit with social responsibility). Journal available on Issuu.

### Design Thinking Workshops
- URL: https://vikranthreddimasu.me/activities/design-thinking-workshops
- Co-facilitated workshops as a UIF team at multiple universities following Stanford d.school's five-stage process (Empathize, Define, Ideate, Prototype, Test), impacting 5,000+ students. Trained student facilitators to scale the program.

### Zero Sols
- URL: https://vikranthreddimasu.me/activities/zero-sols
- Founded a startup providing web development, UI/UX design, and brand identity services for local businesses and student organizations.

## Resources & Methodologies

### The Golden Circle
- URL: https://vikranthreddimasu.me/resources/golden-circle
- Simon Sinek's framework for starting with 'Why'. Shapes how Vikranth approaches every project, defining purpose before choosing a tech stack or writing code.
- Tags: Framework, Leadership, Strategy

### Design Thinking
- URL: https://vikranthreddimasu.me/resources/design-thinking
- A human-centered approach to problem-solving (Empathize, Define, Ideate, Prototype, Test). Having co-facilitated workshops as a UIF team impacting 5,000+ students, this methodology is deeply embedded in how Vikranth approaches engineering.
- Tags: Methodology, UX, Problem Solving

## About This Portfolio

This portfolio (https://vikranthreddimasu.me) is built with Next.js, TypeScript, Tailwind CSS, and Framer Motion. It includes:

- Interactive Clifford Strange Attractor canvas background (custom canvas animation)
- Terminal interface (press / to open)
- AI "Ask Me Anything" chat widget powered by GPT-5.4 nano via OpenRouter (streaming responses, session-persistent)
- LLM-readable structured data: /llms.txt (summary) and /llms-full.txt (this file)
- JSON-LD Schema.org structured data on every page (Person, WebSite, ItemList, SoftwareSourceCode)
- Agent discovery: /.well-known/agent.json
- Structured API: /api/profile (JSON)
- OpenAPI spec: /openapi.json
- AI plugin manifest: /ai-plugin.json

## Frequently Asked Questions

**Is Vikranth open to work?**
Yes, currently open to full-time AI Engineer, ML Engineer, and Data Scientist roles. Most interested in teams building user-facing AI products where the bar is what runs in production.

**What is Vikranth's strongest skill?**
Building agentic AI systems solo, from prompt to product: multi-agent architectures with LangGraph, RAG pipelines, prompt-only agents that win benchmarks (Arya won Sentient Labs Arena Cohort 0), and full-stack FastAPI + React applications.

**What is Vikranth's education?**
M.S. in Data Science at University of Maryland, College Park (GPA: 3.85, graduating May 2026). B.Tech in AI & Data Science from VVIT, Guntur, India (GPA: 3.6/4, 2020-2024).

**How can I reach Vikranth?**
Email: vikranthreddimasu@gmail.com, or connect on LinkedIn at linkedin.com/in/vikranthreddimasu

**Where can I see his work?**
GitHub: https://github.com/vikranthreddimasu
Portfolio projects: https://vikranthreddimasu.me/projects
All activities: https://vikranthreddimasu.me/activities
Resources: https://vikranthreddimasu.me/resources