# Anurag Akkiraju

**Software Development Engineer II** — Seattle, WA

> AI systems engineer building long-horizon agent runtimes and LLM platforms at Amazon, with expertise in prompt caching, multi-agent pipelines, and evaluation harnesses.

Career Archetype: **Frontier Artisan** (RS-MX) — One of the first Frontier Artisans on Saywise

## Links

- LinkedIn: https://linkedin.com/in/anurag-akkiraju
- GitHub: https://github.com/maskedband1t

## Experience

### Software Development Engineer II, Amazon (2022-08 – 2024-07)
Seattle, WA · Building next-generation personal AI agents, owning the runtime for long-horizon durable work and the proactive layer that listens for real-world changes. Led the agent's prompt-caching program, improving cache hit rate from 45.6% to 83.0% while reducing latency by 57%. Led the Alexa+ notification domain rebuild, designing the model interface and evaluation harness for ~90M customers. Designed a platform primitive for deterministic service actions and contributed to the real-time proactive content system, which achieved 26x engagement improvement.

### Machine Learning Research Intern, Lawrence Livermore National Laboratory (2021-06 – 2021-08)
Livermore, CA · Designed and trained an LSTM in PyTorch for sequence labeling over raw binary data, classifying non-code byte regions in PE and ELF binaries at 95%+ accuracy. Built an approximate nearest-neighbor retrieval system for software origin tracing using NMSLIB, returning k-nearest neighbors for newly ingested binaries against an enterprise corpus. Extended the work through spring 2022.

## Education

### B.S. Computer Science, Minor in Electrical Engineering, University of Florida, Herbert Wertheim College of Engineering

## Projects

### Calibrated Decisions at the Human–Robot Boundary
Decision layer between robot planner and policy · Investigated the decision layer between a robot's planner and its policy (act, ask, or hand off to a person). Tested calibrated models against frozen rules and an oracle on four simulated bodies across 160+ experiments. Distilled a 421M-parameter on-device model from a cloud judge, maintaining 88.3% accuracy versus 87.9% for the cloud version at 90ms latency with calibration intact. Identified a failure mode in fleet learning where retraining on operator takeovers can make models confidently wrong outside corrected regions.

## Skills

long-horizon agent runtimes, agent memory, multi-agent LLM pipelines, prompt caching, PyTorch, reinforcement learning, calibration, hybrid FTS5/dense retrieval, evaluation-harness design, instruction specs, frozen baselines, regression testing, non-inferiority testing, failure-mode analysis, Python, Java, C++, TypeScript, SQL, Bash, AWS, Docker, Git, CI/CD, distributed systems, event-driven pipelines

---
Source: https://saywise.com/member6132 (last modified 2026-10-05T19:47:07.521Z)