AI Bottleneck Monitoring
Website and UI still in design
I check this multiple times a day: what's heating up in my themes, what's anomalous on the watchlist, what's changed since yesterday. On-demand monitoring, not a morning briefing.
The problem
AI Bottleneck Monitoring is a monitoring framework with one job: catch supply-chain dislocations before the financial markets price them in, while there's still an edge to act on. The signals that move asset prices, a capex commitment, a capacity constraint, a lead-time blowing out, show up in the physical economy months before they reach an earnings call or a share price. The problem is reach: those signals are spread across a hundred-plus regulatory feeds, exchange filings, and trade-press sources in a dozen languages, and they lose value by the day. No person can watch that surface fast enough. By the time a dislocation is obvious, the edge is already gone.
The approach
- 01 Map the physical layer
The system reasons over a hand-built supply-chain graph of the real economy: 100+ tracked positions across AI, space, commodities, grid, and robotics, wired together by real supplier dependencies.
- One chain: Nvidia designs the chips → TSMC fabs them → ASML, Applied Materials and Lam build the machines those fabs run on → VAT Group makes the vacuum valves inside those machines.
- It scrapes multilingual earnings transcripts, regulatory filings, and trade-press reports across AI and robotics.
- Claude reads each source under a structured prompt and returns physical facts alone: capex commitments, component lead-times, each tagged by type and checked against the registry of tracked companies.
- It pins every fact to the dependency map, so one supplier's constraint lights up everyone downstream and shows who holds pricing power.
- 02 Research it with a four-stage agent
A reasoning agent runs on top of the graph and investigates a single fact end to end. Each of its four stages uses a different model.
- The planner works out which tracked items the fact touches and drafts the research plan.
- The ground stage walks the supply-chain links out from those items and pulls in entities the graph already tracks.
- Workers run in parallel, each taking one slice of the plan. Gaps surface here: a worker on MLCC raw materials finds BaTiO₃ powder cited with no matching node in the graph.
- A final model merges the workers' findings into one report, ranked by whether each item is already tracked, worth adding, or a priority.
- 03 Turn a thesis into who-wins / who-loses
The newest layer, in active development, is a proactive alpha agent. Ask it to 'map every layer of Nvidia's 800V DC power architecture, and every company that supplies into it', and it returns a ranked list of who wins and who loses, graded by exposure.
- The shift runs down a long physical chain: power semiconductors, solid-state transformers, DC-DC converters, busbars, capacitors, cooling. Each layer's suppliers carry different exposure.
- It runs the same shape: the planner breaks the shift into angles → ground anchors each angle to the power-electronics chain in the graph → five workers run in parallel, one per slice → synthesis merges them into one who-wins / who-loses map.
- It stays grounded: every claim ties to the curated graph of real supplier relationships, so its reasoning about who feeds an 800V rack comes from tracked data.
- 04 Keep it honest with evals and AI ops
Evals keep the pipeline honest. They track relevancy and coverage, and they hold the running cost down.
- Relevancy evals score whether the articles the pipeline pulls in stay on-thesis.
- Coverage evals measure the capture rate per sector.
- A curated benchmark refines the extraction prompt for the facts that matter.
In action
How it propagates
Engineering challenges
- Heterogeneous ingestion
The framework only works if every source ends up comparable, because one anomaly engine has to run across all of them.
- The inputs share no common structure: an SEC filing is structured XML, an Asian exchange disclosure is a PDF, FRED is a numeric series, trade press is prose, a news alert is a bare headline.
- They span a hundred-plus sources and a dozen languages, and share one thing: any of them might carry a price, capacity, capex, or lead-time signal worth catching.
- Ingest each in its native shape and you don't have a system. You have a hundred parsers no one can reason about together.
Every source collapses to one structured object at the door. Nothing keeps its native shape.
- Claude reads each raw item and extracts signals, each tagged with one of a small fixed set of types (capex, capacity, lead-time, price).
- The model reports what a source observed, never an interpretation. Observations stay auditable.
- Once every source speaks one vocabulary, a single graph and a single anomaly engine run across all of them.
- Multilingual integrity
The earliest physical-layer signals surface in non-English sources, Japanese, Korean, and Taiwanese filings, because that is where the supply chain sits.
- A Samsung or SK Hynix disclosure can lead an English-language earnings call by months.
- Translation is a silent failure mode. A mistranslation throws no error. It produces fluent, confident, wrong text.
- Because every input becomes a signal on the graph, one bad translation attaches to the wrong position and propagates downstream, a false signal with nothing to catch it.
Translation is an untrusted boundary. Every non-English source has to earn trust before it reaches the graph.
- A verification step checks each translation before the graph ever sees it.
- The system rejects an unverifiable translation at the door, before it becomes noise deep in the graph.
- Propagation without false confidence
The graph earns its value by propagation, and propagation carries a dangerous default.
- A signal on one supply-chain position is a signal on everything downstream of it. That is the point of the graph.
- If a signal crosses a dependency edge unchanged, its certainty crosses unchanged too.
- A weak, unconfirmed signal walked several steps out would arrive looking identical to a confirmed fact, so the graph would manufacture false confidence as a side effect of doing its job.
Confidence fades with distance.
- A derived signal is weaker than a direct observation, and it loses weight the further it travels.
- A signal cannot travel far enough to accumulate false weight.
- The graph still surfaces second-order connections, but a far-away signal can never impersonate a confirmed event.
- Measuring surprise
An anomaly engine that flags every large move is useless, because most large moves are expected.
- Supply chains move all the time: seasonal demand, capex cycles, prices that drift. Flag all of it and the real dislocations drown in noise.
- Size is the wrong test. A price move the framework predicted is not news. A small move that contradicts a confirmed thesis is.
- The hard problem is measuring a signal against what the framework believes, not against zero.
Each signal's weight comes from the thesis it lands on, not from its size alone.
- The framework holds a thesis on each part of the supply chain, and each thesis has a lifecycle state.
- A signal that only confirms what the framework expects adds little. A signal that threatens a confirmed thesis raises a flag.
- The engine flags surprise relative to the current belief, which is what a dislocation is.