Bhanu Pallakonda

Seattle, WA · first AI hire

I own the AI stack for Armada's edge platform — model post-training and agent-harness engineering, from NVIDIA H100/A100 GPUs down to CPU-only nodes. The work has moved through three generations:

Conversational AI assistant

Built the platform's core assistant from scratch on open-weight Llama models with an LLM-as-router design; deployed end-to-end on AKS with vLLM, and quantized for on-device edge inference.

Multi-agent systems

Designed a production supervisor–sub-agent system with modular inference routing; hardened against prompt injection, with observability and evaluation built in.

Skill-agent harness & coding agents

Re-architected the assistant as a composable skill + coding agent — deep agents, code sandboxing, and Armada-specific asset skills — alongside AI-safety research on backdoors in tool-using LLMs.

BERT-based banking chatbot (intent and entity extraction), ID-card detection, and privacy-preserving field-masking pipelines in production.

Pancreatic-cancer prediction on clinical datasets; contract field-extraction with fine-tuned LayoutLMv3.

Armada's core assistant, built from scratch on open-weight Llama models behind an LLM-as-router intent layer. Serving is tiered: vLLM on H100/A100 nodes, quantized GGUF weights via llama.cpp on CPU-only and disconnected sites — deployed end-to-end on AKS.

Llama 2/3 · vLLM · llama.cpp · AKS · FastAPI

A supervisor–sub-agent architecture for task decomposition, routing between reasoning and non-reasoning models across API and self-hosted endpoints. Hardened against prompt injection; observed with Langfuse tracing, an online accuracy evaluator, Grafana alerting, and CI-integrated agentic regression tests.

LangChain / LangGraph · vLLM · Langfuse · Grafana · Redis · Celery

The current agentic layer: deep agents with code sandboxing and Armada-specific asset skills. The assistant plans, writes, and executes sandboxed code against domain assets as a composable skill + coding agent.

deep agents · code sandboxing · tool use · Python

A production embeddings microservice powering retrieval-augmented generation across the platform — vector embeddings and semantic search that ground the assistant and agents in domain data.

embeddings · RAG · vector search · FastAPI · Docker

B. Pallakonda, M. Hindsbo, S. Ehsani, P. Mishra — 2026, first author

Temporal backdoors in tool-using LLMs: latent triggers that surface malicious behavior only under specific conditions, studied through RL fine-tuning and distributed multi-GPU experiments. To be submitted to NeurIPS 2026.

A. Buynitsky, S. Ehsani, B. Pallakonda, P. Mishra — IEEE ICCAR 2025

OPUS, an LLM framework for PTZ camera control with contextual scene understanding — +35% over advanced prompting and +20% task accuracy against closed-source models, via supervised fine-tuning on synthetic data.

S. Ehsani, B. Pallakonda, P. Mishra — ACM SIGSPATIAL 2025

A geospatial analysis of how severe weather degrades LEO satellite-network performance, informing connectivity resilience for remote infrastructure.

Lead patent of the family.