Building deterministic AI systems, not demos.

Software engineer focused on agentic AI systems.

I build:

• multi-agent systems with structured orchestration
• deterministic harnesses for reliability
• evaluation-aware AI systems (focused on correctness, not just output)

Focus areas:

• agent orchestration (LangChain, LangGraph, custom harnesses)
• context + tool design (ACI-style systems)
• AI evaluation (failure detection, reproducibility, ranking)

→ AI systems (LLM workflows, agent orchestration, evaluation)
→ Distributed systems + automation tooling

Deterministic multi-agent CI debugging engine.

Most AI CI tools summarize logs.
They fail because CI failures are execution systems, not text problems.

ci-rootcause reconstructs execution and identifies the actual root cause.

• builds a failure graph from CI logs
• detects the first failure (not downstream symptoms)
• analyzes diffs to link code changes to breakages
• ranks root causes using deterministic scoring
• generates evidence-backed fixes (LLM-constrained)
• produces structured outputs: ci-rca.json, ci-rca.md

LLMs are used selectively for:

• explanation
• fix suggestions

Supported providers:

• Ollama (local models)
• OpenAI
• Anthropic
• Google Gemini

LLMs are never used for scoring or confidence.

• no hallucinated root causes
• reproducible outputs across runs
• confidence is computed, not generated
• works with both local and hosted models
• designed for real CI workflows, not demos

• determinism over heuristics
• systems over prompts
• evaluation before optimization
• evidence over plausibility

• ci-rca.json → machine-readable root cause
• ci-rca.md → human-readable explanation

• contributing to real AI systems (not toy projects)
• building production-grade agent workflows
• focusing on correctness, determinism, and evaluation
• shipping systems that can be reasoned about and verified

Approach: → consistent, high-frequency contributions → focus on real issues that get merged

• [Open] #6535 fix(ethereum): handle trace_filter traces missing result.output via c… in graphprotocol/graph-node
• [Open] #2331 fix(langgraph): handle null thread checkpoint in RemoteGraph.getState in langchain-ai/langgraphjs
• [Open] #5461 fix(converter): fall back on invalid JSON-like partial matches in crewAIInc/crewAI
• [Open] #5545 fix(flow,task): handle pydantic outputs in guardrail retries and checkpoint serialization in crewAIInc/crewAI
• [Open] #2316 fix(sdk): Backfill truncated history for regenerate branching in langchain-ai/langgraphjs
• [Open] #21386 fix(azureaisearch): preserve falsy metadata values in index mapping in run-llama/llama_index
• [Open] #21336 fix(elasticsearch): split sync and async store paths in run-llama/llama_index

• Determinism over heuristics where possible
• Systems over prompts
• Evaluation before optimization
• Evidence > plausibility

• LangChain
• LangGraph
• multi-agent orchestration
• LLM tool + context design

• Python
• TypeScript
• Rust

• agent harness design
• evaluation systems
• reproducibility + reliability

• GitHub
• LinkedIn