Brand Concierge (BCOS) — Enterprise Multi-Tenant LLM Agent Orchestrator, Grounded Neural Retrieval & Automated Program Optimization at Adobe
- Agent Orchestration Engine: Built the Brand Concierge Operating System (BCOS) with PydanticAI, declarative manifest compilation, and Agent-to-Agent JSON-RPC support.
- Declarative Tool Framework: Designed extensible Pydantic tool schemas with declarative pre/post execution hooks, state capture primitives, and configurable
ToolErrorPolicyfor deterministic runtime recovery. - Neural Source Attribution: Integrated remote cross-encoder rerankers (bge-reranker-v2-m3, e5-curriculum) with score normalization, token limits, and request batching.
- High-Performance Inference (Lucid Lens): Built CPU-only Text Embeddings Inference services on ONNX Runtime with dynamic axes for batch reranking of 30+ passages, with reported sub-second p50 latency.
- Automated Agent Optimization: Integrated GEPA and DSPy into Eval Studio for reflective prompt compilation.
- Streaming Serving & Traffic Isolation: Hardened core FastAPI/MongoDB conversation engine with Server-Sent Events (SSE) token streaming, artifact state replacement (
StreamEventProcessor), and sandbox isolation (TrafficControlEnforcer).
- Evaluated across four causal language model families on numerical reasoning (GSM8K, SVAMP), reading comprehension (DROP), and logical reasoning tasks.
- Improves Zero-Shot-CoT accuracy on
code-davinci-002by 5% on numerical reasoning and 13% on reading comprehension. - Demonstrates that both the original query and the model-generated rephrased version provide complementary semantic signals that stabilize in-context learning.
Online visual interface for exploring the effect of pretraining term frequencies (The Pile) on few-shot language model performance across GPT-Neo, GPT-J, OPT, and BLOOM.
Investigating activation properties and feature representations of deep neural networks during Kharagpur Winter of Code (KWoC).
React implementation of the 2048 sliding tile puzzle game.
Updating PageRank scores incrementally for web pages in large, dynamically evolving web graphs without full power-iteration recomputation.
Stan Weinstein Stage 2 momentum scanner, market intelligence reports, and deep business monographs.
Structured destination guides, itineraries, and field research.