I build retrieval pipelines, agent workflows and MCP tools.
With evaluation and observability built in.
SELECT A NODE TO INSPECT THE FLOW
Three ways to make complex AI systems
easier to understand—and work with.
One place to connect to MCP servers, discover their tools and see exactly what a request returns.
One exchange, laid open.
Select a method below to inspect its shape.
{ "jsonrpc": "2.0", "method": "tools/list" }{ "tools": [{ "name": "lookup", "inputSchema": { … } }] }Debug the exchange. Inspect the request, run the tool, follow the response.
Connect / discover / execute / inspect
Route the task. Coordinate the agents. Follow every handoff from the first plan to the final tool call.
Every agent has a role.
Every handoff leaves a trace.
task.received←topology.created·schema.invalid·patch.started·execution.complete·A multi-agent workflow needs more than an answer. It needs a trace you can follow.
Explicit handoffs / execution state / live events
Classify the message, retrieve its context and draft the response. Route uncertain decisions to a person before dispatch.
The message pauses at the review gate.
SIMULATED CONFIDENCEA message arrives. The system finds context before it acts.
The context leaves room for doubt.
Pause here for a person’s decision.
The draft clears the routing gate.
Continue toward dispatch.
Human review is a route through the system—not a workaround when the model gets stuck.
Context retrieval / response drafting / review gate
What happens before
the model ever answers?
The agreement defines the services, obligations and responsibilities of each party. Defined terms apply throughout this document.
Either party may terminate this agreement upon thirty days’ written notice. Obligations incurred before termination remain in effect.
This agreement shall be governed by and interpreted in accordance with the laws of the designated jurisdiction, without regard to conflicts of law.
REFERENCE SPANEach party shall protect confidential information received from the other party and use it only for the purposes of this agreement.
Fixed-width boundaries split reference spans. The answer is broken before the model sees it.
Published results from the retrieval evaluation harness, not a live benchmark of this illustration. Span integrity measures chunk boundaries—not answer accuracy. Small sample; no confidence intervals.
READ THE EXPERIMENT ↗INSPECT THE HARNESS ↗Before another rebuild, find out what your retrieval actually gets wrong. A focused diagnostic of the system you already have.
40 labeled queries with relevance judgments, approved by you. Each answer is located by character offset in your own documents, so relevance is decided by span arithmetic rather than string matching.
LET’S LOOK AT YOUR PIPELINE ↗ONE EXISTING PIPELINE / UP TO 250 DOCUMENTS OR 2 GB — ENFORCED BY THE HARNESSSystems built with teams.
Decisions made under real constraints.
Building multi-agent workflows and RAG pipelines for enterprise clients. Working on retrieval quality through semantic chunking and cross-encoder reranking.
Converted from the internship that ran Jan – Sep 2026.Semantic chunking / cross-encoder reranking / agent workflowsMake retrieval behavior measurable before adding more orchestration.
From the interface to the infrastructure.
Select a layer to inspect its role.
Interfaces that expose state, explain the next action and give people a way to inspect what happened.
Hackathons and engineering challenges.
A different kind of system test.
AI-driven orbital mechanics and tracking solution built for high-signal decision support under hackathon constraints.
The boundary is already a decision. An experiment in what survives when a contract becomes chunks.
READ THE EXPERIMENT ↗PUBLISHED EXPERIMENT / ILLUSTRATIVE DOCUMENT
Start with a question you can test.
I work across the stack, with a particular interest in retrieval, agent coordination and making complex behavior inspectable.
Measure before optimizing.
Observability is part of the product.
Know when to bring a human in.