Selected public and anonymized systems work. The pattern is consistent: start from a real operational problem, design a clearer workflow, then use software or AI only where it actually helps.
Public
Private
Current
Selected evidence
Turning responsibility, cost pressure, and local logistics friction into something people can inspect instead of memorize.
Responsibility map
seller
seller
handoff
buyer
Editorial scenario—not a freight quote or legal allocation.
Synthetic interface
08
On track
03
Needs review
02
Waiting docs
Synthetic labels and counts. The design shows information hierarchy, not customer or shipment records.
Agent systems · evidence method
Each workflow starts with a task definition, expected output, visible failure case, and a review boundary before any tool can act.
Explore the interactive routing modeldefined
checked
visible
required
Case notes
Problem
Trade responsibility, landed cost, and local logistics friction are hard to explain with rules alone.
Work
A public research tool that turns Incoterms into interactive responsibility maps, cost simulations, and lab notes about supply-chain improvement in Indonesia.
Result
Makes abstract trade terms easier to discuss with scenarios, seller/buyer splits, and operating-environment stress tests.
Problem
Daily operational work was scattered across conversations, files, and legacy records.
Work
A private dashboard for turning shipment and workflow signals into a morning view: what is moving, what is stuck, and what needs attention.
Result
Replaces scattered tabs and chat threads with one morning view operators can act on, while customer- and shipment-level data stays out of the public signal.
Problem
Incoming operational messages need classification, field extraction, and routing before anyone can act on them.
Work
A private intake layer for reading unstructured messages, extracting useful fields, and turning them into clearer follow-up tasks.
Result
Routine messages sort and route automatically; ambiguous or risky cases stay visible to a human reviewer before any action is taken.
Other systems
Each track ships only after small evals on a real operational task — the public signal stays simple, the internal work stays verifiable.
Keeps useful history accessible while reducing the blast radius of legacy-system change.
Turns self-taught learning into reusable judgment before tools are trusted with real work.
Keeps the public story clean: operations, systems thinking, applied AI, and supply-chain context.
Private R&D
The useful question is not small model versus big model. It is which model should handle which job, under which guardrails, with what context, cost, and review boundary.
Local models can be excellent for repeatable structured work. Bigger models are useful when the task needs planning or messy synthesis. The system around them matters more than the model name: routing, memory, tools, evals, and escalation decide whether an agent is actually safe to use.
Decision rule
The router decides local model, large model, or human review based on task shape, context, and risk.
Structure needed before local models are useful
Private work note
Internal systems are described by shape, not by private records, infrastructure details, or operational traces. If the context is useful, I can walk through the thinking in a private conversation without exposing the parts that should stay private.
Contact me about similar work