Skip to content
Aneeq Khatri

demos

The systems, running

Walkthroughs and screens captured from the real applications. Everything here was rendered by the project it belongs to. No mockups, no concept art. Projects that have not been captured yet are simply absent rather than represented by something that looks like a screenshot.

13 of 16 projects have a walkthrough or screens · all data synthetic

LedgerLens

Read the case study

An exception workbench where an uncited hypothesis is deleted in code, applied to the breaks AI matching cannot close

A narrated run through the exception workbench: import, the queue, opening a transaction, and reading the evidence behind a match before confirming it.

1 / 4 · screens

The review queue: 25 exceptions with root-cause chips, confidence pills and a batch progress bar.
The review queue. Root-cause chips, confidence pills, and a batch run in progress over the WebSocket. View full size
An exception detail view in three panes: the transaction, ranked hypotheses with evidence chips, and the evidence panel with matched fields highlighted.
One exception. Ranked hypotheses, each with the evidence chips it cited. Click one and the source record opens with the matching fields highlighted. View full size
The pattern library, grouping confirmed resolutions by root cause and counterparty.
The pattern library. Confirmed resolutions group by root cause and counterparty, so the third occurrence arrives with 'seen 3 times before' attached. View full size
The resolution report, showing the AI top-one agreement tile above an immutable verdict log.
The resolution report. Note the caveat in the limits below: the agreement figure only means something once corrections are in the log. View full size

LedgerLab

Read the case study

An open MCP server that hands any agent a realistically messy world, applied to bank and ledger reconciliation

A nightmare-profile world of 105 transactions, a live GPT-5.6 Luna agent reconciling it over MCP, then the answer key revealed: 52.1% coverage and 100% precision. It matched about half of what it could, and every match it made was correct. That is a live model run; the table below is the deterministic reference heuristic. The presenter is AI-generated; the project and the numbers are real.

1 / 4 · screens

The session detail screen: 64 transactions, 61 matches proposed, 24 exceptions flagged, and a table of the agent's proposed matches with its stated reason for each.
One reference-agent run against a seeded world. Note the line under the progress bar: it is progress, not accuracy. The tools never tell an agent whether a match is correct. Grading happens against a ground-truth table no tool can reach, and the answer key is viewer-only. View full size
The live tool-call feed: each row shows the tool name, its arguments, latency in milliseconds and payload size in bytes.
Every tool call as it happens. Payload size is on each row because that is what the call costs the agent's context window. Which is the reason results are paginated and capped at 50 rows in the first place. View full size
The world browser, showing counts per entity: bank feed, general ledger, invoices, counterparties, notes and chart of accounts.
The generated world, entity by entity, with the messiness annotated. Aliased payees, missing references, chaotic date formats, so you can see what the agent is actually up against. View full size
The sessions screen: a card per world showing profile, seed, transaction count and the dataset hash.
A card per world. The dataset hash is on the card because same profile plus same seed must produce a byte-identical world, here it matches the value committed in CI. View full size

CloseOps

Read the case study

Five governed agents with a provable no-duplicate guarantee, applied to month-end close

1 / 3 · screens

The CloseOps board: seven tasks in four columns by dependency depth, each with risk tier and gate chips. One reconciliation task is red.
A completed close. Seven tasks, three specialists, five gates, and the planted anomaly in red: account 1000, 17 of 17 transactions unmatched against a 55% tolerance, escalated and resolved by a human. View full size
The gate queue, listing tasks awaiting a human decision with their risk tier and the reason each was gated.
The gate queue. It mirrors the Teams card and posts the same signed callback, so the console has no privileged path of its own. View full size
The close summary: time to close, tasks auto-completed versus gated, anomalies caught, and duplicate postings.
The summary the whole project exists to produce: auto versus gated, anomalies caught, escalations, and duplicate postings at zero. View full size

FinAgent-Evals

Read the case study

A vendor-neutral benchmark for reliability under repetition, applied to finance reconciliation tasks

1 / 4 · screens

The leaderboard: two configurations with pass@1 → pass^k columns, a consistency-gap badge on each row, and three disclosure banners above the table.
The board leads with the gap, not the rank. mock holds 94.0% across all four repetitions; mock-flaky drops 45.8% → 8.0%, and the −37.8% badge is the number a single-shot eval would never have shown you. The three banners above the table are not a footnote. Synthetic data, a release-candidate suite, and no live runs are stated before any score is. View full size
The compare view, showing headline, pass@1, pass^4, consistency gap, MatchF1, EscalationScore and PathValidity side by side for two configurations.
The same two configurations metric by metric. pass@1 falls 48.2% while pass^4 falls 86.0%. The spread between those two numbers is the whole argument for running every case k times. View full size
The case explorer filtered to tier-4 traps, with one case open showing its seed, trap family, and a table of bank transactions in four different date formats with no references.
A tier-4 trap, open. Every case here contains a transaction where escalating is the only correct action, and the answer stays behind a Reveal button so you see exactly what the agent sees. Note the dates in the transaction table. 26/03/2025, 04-16-25, 03 Mar 2025, 28.02.25, and the references reading none. View full size
The methodology page, explaining what the benchmark measures and justifying each weight in the headline formula.
Every weight in the formula is argued for on the page rather than asserted. pass^k takes the largest share because an agent that is right once in four is not 25% useful in a month-end close, it is unusable. View full size