# Pragnyan Ramtha | AI Engineer Full Context

AI Engineer specializing in LLM Fine-Tuning (PEFT/QLoRA), Autonomous AI Agent Systems, and Cost-Efficient Model Compression (GPTQ/Mixed-Precision). AIMO-3 Solver Medalist, ARC-AGI #1 Ranker, and Top 5 OpenAI Parameter Golf.

## Contact
- GitHub: https://github.com/pragnyanramtha
- LinkedIn: https://www.linkedin.com/in/pragnyanramtha
- Professional Resume: /resume.pdf
- Site Index: /index.md

## Technical Focus Areas
- **LLM Fine-Tuning:** PEFT, QLoRA, CoT/TiR datasets, Phi-4 optimization, contrastive learning.
- **Autonomous AI Agents:** Agentic loops with parallel tool-use, BFS/DFS decoders, consensus-based voting, long-running orchestration.
- **Model Efficiency:** Mixed-precision quantization (GPTQ Int6), Cross-Sparse Attention (XSA), EMA weight smoothing, extreme compression.
- **Test-Time Adaptation:** Test-Time Training (TTT) strategies for solving novel abstract puzzles (ARC-AGI).
- **RAG Systems:** Retrieval-augmented generation pipelines with semantic search and source attribution.
- **MLOps:** High-availability serverless backends on GCP with aggressive cost-cutting logic.

## Key Projects & Achievements
- **Agent7:** Solo-built no-code AI agent platform connecting 1,000+ app connectors, orchestrating 20,000+ tools, running autonomous sessions for hours to days.
- **AIMO-3 Solver Medalist:** Fine-tuned Phi-4 (14B) to 90% reasoning accuracy on competition benchmarks rivaling 125B parameter models.
- **OpenAI Parameter Golf (Top 5):** Achieved 1.1271 BPB on FineWeb using XSA4 + EMA + GPTQ-Int6 stack.
- **ARC-AGI-2 (#1 Rank):** Adaptation of budget-aware solving patterns and Test-Time Training for fluid intelligence benchmarks.
- **Personality Clone:** Fine-tuned LLM with QLoRA + siamese network architecture achieving 92% style replication accuracy (28% improvement over baseline).
- **Agentic AI Developer:** Built GCP serverless AI orchestration, Gemini API workflows, and RAG systems at Reputation Dao.

## Open Source Contributions (70+ merged PRs)
- openai-node, langgraphjs, pydantic-ai (highlighted)
- promptfoo, haystack, mem0, chainlit, agno, fastmcp, dify-plugins, dstack, opik, mastra
- Gemini CLI (sudo bug fix), Scrapy (duplicate filter reliability patch)

## Research Papers
- **Scaling Context Windows to Infinity** — Analysis of position encoding and memory-efficient inference for long-sequence processing (Academia.edu, 2026).
- **Unlocking Societal Trends in Aadhaar Enrolment** — ML approach to fraud detection in biometric systems (Academia.edu, 2026).
- **Quantum-Inspired LLM Inference** — Investigating variational quantum circuits for token generation acceleration (In Progress).

## Blog Summaries
- **Why OpenAI Sent Me $500 (Parameter Golf):** Technical breakdown of XSA4, EMA, and GPTQ-Int6 for 16MB model training.
- **How I Reached #1 on ARC-AGI-2:** Test-Time Training strategy for fluid intelligence benchmarks.
- **How I Won a Solver Medal at AIMO3:** Parallel agentic loops with sandboxed code execution and budget-aware time allocation.
- **Fine-Tuning Reasoning:** Insights on data quality over quantity and benchmark hygiene.
- **LLM Latency:** Systems-level approach to breaking latency budgets across retrieval and inference.

## Skills
- languages: [Python, TypeScript, Rust, C, SQL, Bash, Go]
- ai/ml: [PyTorch, Transformers, Unsloth, PEFT/QLoRA, CUDA, RAG Pipelines, AI Agents, LLM Fine-tuning]
- full-stack: [Next.js, React, Node.js, Tailwind CSS, PostgreSQL, REST APIs, SaaS Architecture]
- infrastructure: [GCP, Docker, Linux (Arch), Nix, Git, CI/CD]
- developer-tools: [Neovim, Arch Linux, uv]

## Education
- B.E. Computer Science & Design, MRV University (GPA 9.0/10, Top 3 academic rank).
- Machine Learning Specialization (Stanford), CS50 (Harvard).