Pragnyan Ramtha

• Built end-to-end post-training data pipelines for frontier AI labs, curating expert SFT and RLHF datasets across reasoning, tool use, and code synthesis.
• Created and audited evaluation benchmarks and agentic coding tasks for private frontier models, using deterministic execution harnesses, test-based verifiers, and contamination checks to prevent reward hacking.
• Audited multi-turn agent trajectories and verification loops to improve data quality and post-training signal.

• Post-Training
• LLM Evaluation
• Benchmarks
• RLHF
• SFT
• Agent Harnesses

• Shipped a no-code AI agent platform from zero to production solo, connecting 1,000+ app connectors and orchestrating 20,000+ tools by building the full stack: agent engine, approval system, integrations, and no-code frontend.
• Enabled teams to hand off 20+ hours of weekly repeat work, agents run autonomously for hours to days across research, outreach, CRM, and cross-app workflows by designing a long-running orchestration system with approval-first safety controls.

• Next.js
• TypeScript
• AI Agents
• Full-Stack Development
• SaaS

• Built assistive software for blind learners from scratch, shipping a learning portal with screen-reader compatible workflows by engineering accessibility-first interfaces and AI-powered assistance tools.
• Delivered accessible class sessions for visually impaired students, adapting tools and workflows for screen-reader compatibility across real-time sessions by collaborating with educators and TEDx speakers to design inclusive learning experiences.

• Accessibility Engineering
• Accessibility
• Python
• Learning Portals
• AI Assistance

• Architected a GCP serverless backend achieving 99.9% uptime by leveraging Cloud Functions and Cloud Run for production-grade AI orchestration.
• Reduced inference latency by 50% across support workflows by engineering a Gemini API response system with optimized prompt caching.
• Boosted accuracy and trust by developing a RAG pipeline utilizing semantic search for real-time documentation retrieval and source attribution.

• Web Development
• Full-Stack Development
• GCP
• Gemini API
• RAG
• Python

• Researched world-model approaches for architecture workflows to generate CAD-style design outputs from spatial context and architect design intent.
• Prototyped ML pipelines that translated early architectural concepts into structured geometry, creating a faster path from design exploration to CAD handoff.
• Evaluated generated layouts against architectural constraints to improve reliability before model outputs were used in downstream design workflows.

• Artificial Intelligence
• World Models
• CAD Generation
• Machine Learning
• Python

• Designed and released an open benchmark of 212 everyday writing tasks, run on 11 models with three samples per task. The leaderboard, the paper, the task cards and every public answer are published.
• An output counts only if it passes three gates: programmatic checks, a cross-lab LLM judge, and an AI-text detector calibrated on 801 human-written texts. The top model passes all three on 23.1% of outputs. Scores are reported on public and held-out tasks with 95% bootstrap intervals.

• Python
• LLM Evaluation
• Benchmark Design
• LLM-as-Judge
• AI-Text Detection

• Built a no-code AI agent platform from scratch, solo. Agents connect to 1,000+ apps and run autonomous sessions for hours or days.
• Handles research, outreach, CRM updates, lead qualification, and cross-app workflows from plain English instructions.

• Next.js
• TypeScript
• AI Agents
• Integrations
• SaaS

• Open-source CLI and daemon that lets AI coding agents message each other: Claude Code, Codex, OpenCode, Cursor, Gemini CLI and others, on one machine, across machines, and across a team.
• Delivers messages into running sessions, wakes idle agents, and shows who is busy or offline. Team messages are end-to-end encrypted per device, so the relay only sees ciphertext. Ships with loop guards, rate limits and a secret scanner.

• TypeScript
• Node.js
• MCP
• Agent Tooling
• E2E Encryption

HumanBench: Can AI Write Like a Human? 2026

Read paper Research preview v0.1

A benchmark of 212 everyday writing tasks, run on 11 models with three samples each. An output counts only if it passes programmatic checks, a cross-lab LLM judge, and an open AI-text detector calibrated on 801 human-written texts. The best model passes all three on 23.1% of outputs. The paper reports public and held-out scores with bootstrap intervals and sets out the benchmark's limitations.

Scaling Context Windows to Infinity: A Comprehensive Study of Position Encoding, Attention Mechanisms, Memory-Efficient Inference, and Context Reduction Techniques in Large Language Models 2026

Read paper Academia.edu

A comprehensive analysis of techniques for extending context windows in large language models, examining position encoding strategies, efficient attention mechanisms, and memory-optimized inference approaches to enable processing of arbitrarily long sequences.

Unlocking Societal Trends in Aadhaar Enrolment and Updates: Anomaly Detection and Fraud Risk Prediction 2026

Read paper Academia.edu

A data-driven approach to identify suspicious patterns in India's Aadhaar biometric identification system, utilizing machine learning for anomaly detection and fraud risk prediction in enrollment and update processes.

Speeding Up LLM Inference Using Quantum Computing Techniques 2026

In progress

Exploring quantum-inspired algorithms and quantum computational primitives to accelerate inference in large language models, investigating quantum annealing for attention mechanisms and variational quantum circuits for efficient token generation.