work
a collection of AI systems i've built, from memory infrastructure and retrieval research to fine-tuned language models and enterprise agent platforms.
- Persistent memory infrastructure for AI agents, scoring 89.0% task-averaged accuracy on LongMemEval-S (84.4% overall, 500-question long-term memory benchmark), evaluated with an LLM-as-judge Gemini reader
- Retrieval pipeline combining semantic, temporal, BM25 keyword, and reranking channels, merged with Reciprocal Rank Fusion and re-weighted by a rule layer that learns from outcomes logged on prior queries
- CABAR, a context-budget-aware retrieval pipeline for long-context agents: greedy token allocation, lexical MMR, recency weighting, and extractive summarization — cutting context usage by ~72-76% in internal evaluations at equal benchmark accuracy
TypeScriptPostgreSQL (pgvector)RedisNeo4jOpenAIGemini
DRT — Decomposed Relevance Tensors
- Engineered and benchmarked an alternative retrieval-scoring architecture — splitting each embedding into k=6 weighted sub-vectors instead of one flat similarity score — against a cosine-similarity baseline trained on the full MS MARCO dataset (8.84M documents, 6,980 queries)
- Ran a 4-step diagnostic (scoring ablations, per-slot logistic probes, failure analysis, linear CKA) to isolate why it underperformed baseline by 2% MRR@10, tracing the cause to encoder representation drift (CKA 0.945) rather than a broken scoring function
- Open-sourced the full pipeline — training, evaluation, and diagnostics — with all raw results and logs checked in for reproducibility
PythonPyTorchMS MARCO
- Coding LLM fine-tuned specifically for Solana/Anchor development: QLoRA fine-tune of Qwen2.5-Coder-7B on 270k Solana-specific records
- 85% on an 80-task Solana/Anchor benchmark, 90% on general HumanEval
- RAG pipeline over current Solana docs (Qdrant) keeps it accurate as Anchor syntax evolves
- Served via an OpenAI-compatible API across web, CLI, VS Code, and MCP, with direct integration into Claude Code, Cursor, and Windsurf
- 1,500+ installs on PyPI
PythonPyTorchSGLangQdrantMCPNext.js
- Enterprise agentic AI platform that orchestrates workflows across 17 live app clients (ServiceNow, AWS, Jira, Salesforce, SAP, Snowflake, SharePoint) via a function-calling loop where Gemini selects and chains real API actions, with progress streamed over SSE
- Enforces role-based access control at query time, filtering RAG results against each user's designation before content reaches the LLM, so employees only retrieve answers from HR, IT, CRM, and compliance sources they're authorized to see
- Multi-step workflow engine that sequences AI agents with variable passing between steps, mid-workflow file-upload pausing, and automatic email notification on completion
FastAPINext.jsGoogle GeminipgvectorPostgreSQLAWS