Prime Agent Explored
Aug 6, 2026
Agents | Inference Engineering | Retrieval | Memory | Observability | Evals
Side Quests
Helios
Helios
Lightweight but powerful Python inference engine for Qwen3-4B on CUDA GPUs, with KV caching, prefix caching, paged attention, continuous batching, and disaggregated prefill/decode.
Ares
Ares
Self-improving coding agent with a persistent Python workspace, isolated subagents, and durable local sessions.
Evalon
Evalon
Local, terminal-first observability and evals for Python agents. Traces, versioned datasets, no cloud.
Atlas
Atlas
Semantic memory with a 3D brain atlas and MCP tools — typed, deduplicated memories by meaning.
Writing
Notes from the build.
Experience
AI systems I have helped ship.
Binoloop
AI Engineer
Apr 2025 – Dec 2025
- •Shipped a production RAG system with BM25, dense retrieval, query expansion, reranking, and conversational memory, improving retrieval relevance by 40% and user engagement by 25%.
- •Created an LLM prompt-optimization pipeline with chain-of-thought and few-shot examples, improving regulation and urban-planning document analysis accuracy by 35%.
- •Cut response latency by 78% and memory footprint by 71% via connection pooling, leaner data structures & pipeline refactoring.
- •Automated evaluation with an LLM-as-a-Judge framework using GPT-4.1, reducing manual annotation effort by 80%.
Lamarr
Artificial Intelligence Intern
Sep 2024 – Mar 2025
- •Built FastAPI services to automate scraping of 10,000+ Indian court case files, improving legal data accessibility by 60%.
- •Designed a LangGraph multi-agent system for contract analysis, identifying adversarial clauses with 85% precision.
- •Delivered a production RAG pipeline with hybrid search for a legal Q&A chatbot, improving answer relevance by 40%.
- •Implemented document preprocessing and chunking workflows for long-form legal text, improving retrieval consistency and reducing noisy context in downstream contract and case-law analysis.
Findability Sciences
Data Science Intern
Jan 2024 – Jun 2024
- •Built an end-to-end time-series forecasting pipeline with XGBoost and feature engineering, achieving 98.8% accuracy on a 10,000+ row dataset.
- •Developed hierarchical classification datasets across five business domains using custom ontologies.
- •Created model evaluation dashboards and error-analysis reports to compare forecast performance across segments.
Activity
GitHub and toolchain.
Stack I use
Technologies I work with to build AI systems and applications
Python
SQL
Bash
PyTorch
FastAPI
Pydantic
LangChain
LangGraph
LlamaIndex
n8n
Milvus
GCP
Docker
Kubernetes
Supabase
PostgreSQL
SQLite
Redis
Langfuse
Postman
Git
GitHub
MCP
Python
SQL
Bash
PyTorch
FastAPI
Pydantic
LangChain
LangGraph
LlamaIndex
n8n
Milvus
GCP
Docker
Kubernetes
Supabase
PostgreSQL
SQLite
Redis
Langfuse
Postman
Git
GitHub
MCP
Python
SQL
Bash
PyTorch
FastAPI
Pydantic
LangChain
LangGraph
LlamaIndex
n8n
Milvus
GCP
Docker
Kubernetes
Supabase
PostgreSQL
SQLite
Redis
Langfuse
Postman
Git
GitHub
MCP
Python
SQL
Bash
PyTorch
FastAPI
Pydantic
LangChain
LangGraph
LlamaIndex
n8n
Milvus
GCP
Docker
Kubernetes
Supabase
PostgreSQL
SQLite
Redis
Langfuse
Postman
Git
GitHub
MCP