Navicade Labs navicade.com

We build the software small businesses actually run on.

Dashboards, portals, internal tools, the operational systems behind the day.

See the work

Work

Bordero, the commission reconciliation desk Commission reconciliation Bordero Statements matched line by line, clawbacks to the cent. View project → The Ledgerline app, a real screenshot Payments console Ledgerline Volume, risk scores, a fraud review queue. View project → The Hartwell & Vane app, a real screenshot Practice platform Hartwell & Vane Matters, billable timer, trust accounting, intake. View project → The Traceline app, a real screenshot AI observability Traceline Traces, token streams and evals for agents. View project → The Meridian Issues app, a real screenshot Issue tracker Meridian Issues Board, list, burndown, and a command palette. View project → The Northbeam app, a real screenshot Product analytics Northbeam Funnels you edit, cohorts, live segmentation. View project →

Capability demo

One build kept purely as proof of engineering under load. It is a market simulation and says so on its page; nothing about it claims client work.

The Vantage Markets app, a real screenshot Trading terminal Vantage Markets Live candles, a real order book, orders that fill. View project →

How we work

01

Audit

We map the operation end to end, not a feature list. Where the money leaks, where the hours go, and what breaks the week someone is on holiday. You get the findings whether or not you build with us.

02

Pilot

We build the first fix and put it in front of the people who have to use it. Narrow enough to prove in weeks, real enough to keep. If it does not earn its place, it stops here.

03

Retainer

We run the system with you, ongoing. Not a handover and a goodbye: software that nobody owns is software that quietly rots, so we stay and keep it working as the business changes.

A portfolio of working builds. Open any one and use it. The names and numbers inside are mocked.

No requirements document, no discovery theatre. You see something working in weeks, not a deck in month three. Every layer we ship is usable on its own, so the work earns its keep before the next piece arrives.

← All the work

Commission reconciliation

Bordero

Built for: Insurance agencies and advisory firms that find out what they were actually paid one statement at a time.

The problem

Every month the insurer sends a bordero, and every month somebody in the office checks it against what the firm actually wrote: line by line, rate by rate, by hand. The mismatches are small, constant and real money: a rate paid short, a cancelled policy still paid, a clawback nobody can trace. Most firms find out what they earned by trusting the statement.

What we built

The match, live. Open a 52-row statement and run it against the production register. Rows check off one by one; five land as exceptions, each carrying its own arithmetic: agreed 15% of €1,240.00 is €186.00, paid €148.80, short €37.20.

A queue with decisions. Approve posts the difference to the ledger. Dispute files it with the insurer on a visible 90-day clock. Both land on the record with a name attached.

Two markets, one desk. Flip Athens to Singapore and the same screens hold: EUR becomes SGD, premium tax becomes GST on the firm’s own fees, producer splits become adviser hierarchies with a supervisor override, and the first-year cap-and-spread schedule computes live, clawback included.

Roles that mean something. A principal sees the firm; a producer sees only their own book. One switch, honest consequences.

Decisions we made

The rule the desk defends: commission is earned on collected premium, not written premium, and unearned commission is never the firm’s money. Athens shows it as accrued against payable; Singapore shows it as the spread schedule and a clawback that reverses exactly the unearned part, from everyone who was paid.

Insurers, firms and people here are invented, and the market conventions are the common ones, not a compliance claim. The mechanics are the point.

On Monday morning

The statement check that ate an afternoon becomes a run button and a short queue, and the argument with the insurer starts from a filed dispute with arithmetic attached instead of a feeling that the number looks low.

What to try in the demo

  1. Open the Delphi Mutual statement and run the match
  2. Open the rate-variance exception, check its arithmetic, dispute it
  3. Add your own policy to the register, run the match again, watch it get flagged as missing
  4. Switch to the Singapore workspace and lapse the year-two policy to watch the clawback cascade
  5. Switch the role to Producer and watch the book contract
george@navicade.comTell us what you are building. We answer within one business day.
← All the work

Capability demo · trading terminal

Vantage Markets

Built for: Anyone who needs to believe we can build the hardest-looking software there is.

The problem

Trading software is the stress test of interface engineering: hundreds of numbers changing every second, charts that must redraw without stutter, and money math that has to be exact while everything moves. Most portfolios show a screenshot of one. This one runs.

What we built

A candlestick chart drawn from scratch. Canvas, ninety bars, volume underneath, a crosshair that reads out OHLC as your finger moves, three timeframes that regenerate the series.

An order book with consequences. Ten levels each side that persist and deplete: a market order walks the displayed levels and fills at the volume-weighted price, so size visibly moves the price you get.

Limits that rest and a desk that says no. A limit under the market sits working until price comes to it, cancel returns the reserved buying power, and an order the desk cannot afford is rejected, not absorbed.

Decisions we made

Prices come from a random walk with volatility clustering, the same math real simulators use, so the chart moves like a market instead of sitting still.

Invented tickers by design. Real ones would imply real data we do not have.

This page sits under its own heading because it is the flex, openly. Prices are a simulation. The engineering is not: book, fills, buying power and profit math are one system, and you can try to break it.

On Monday morning

This is the reference app: if the studio can keep sixty numbers coherent at 800 millisecond ticks on a phone, the dashboard your business actually needs is a quiet afternoon.

What to try in the demo

  1. Hover the chart and flip 1M to 1D
  2. Open BOOK and watch the tape print
  3. Buy 100 AXOM, wait for the fill, then find it in positions repricing live
  4. Try to buy 5,000 AXOM and meet the rejection: buying power is enforced
  5. Switch the ticket to LIMIT, rest an order under the market, watch it fill when price dips
george@navicade.comTell us what you are building. We answer within one business day.
← All the work

AI observability

Traceline

Built for: Any company putting an AI agent in front of customers and wondering what it actually does all day.

The problem

An AI agent in production is a black box that spends money and talks to your customers. When it fails, you need to see exactly which step failed and why; when you improve it, you need proof the new prompt is actually better. That tooling barely exists, and we build it.

What we built

Every run, dissected. A waterfall of spans: agent, retriever, model calls, tool calls, each with its timing bar and its actual payload, expandable.

Failure made visible. One trace shows a rate-limited tool call in red, and the retry that saved it. Error handling is the part that proves production experience.

Evals, not vibes. Prompt V3 against V4 across six criteria and 120 replayed conversations, with the sample answers side by side. The rollout decision becomes evidence.

Decisions we made

Costs shown at real precision and latencies log-normal, because anyone who runs agents knows what those distributions look like.

The replay streams token by token, the way the model actually answered.

Every trace renders from its own run’s facts. Open two runs and the plans, tool calls and answers differ, because an observability tool that tells the same story twice is decoration.

On Monday morning

When your agent tells a customer something wrong, you find the exact span that caused it before lunch, and the eval table tells you whether the fix helped or just felt better.

What to try in the demo

  1. Open the run marked ERROR and find the retry
  2. Replay the answer stream on an llm span
  3. Open two different runs and compare their plans and answers
  4. Flip V3 to V4 in evals and read what actually improved
george@navicade.comTell us what you are building. We answer within one business day.
← All the work

Issue tracker

Meridian Issues

Built for: Software and product teams who live in their tracker eight hours a day.

The problem

Issue trackers are where software teams live, and the good ones feel instant: keyboard-first, zero loading states, every mutation everywhere at once. Clients cite Linear by name. Matching that feel is the test.

What we built

One array of truth. Twenty-six issues drive board, list, counts and analytics; move one anywhere and everything repaints together.

The palette. Cmd K opens a command bar that executes sixteen real commands, assign, reprioritize, move to Done, jump views, and finds real issues by title: type billing and it takes you to the proration ticket. With a toast to prove it happened.

A cycle that reads honest. Burndown with ideal against actual, a scope-change annotation, and velocity compared to last cycle.

Decisions we made

Issue titles are real engineering sentences, comment counts follow a power law, and timestamps cluster in working hours, because reviewers check exactly these things.

On Monday morning

Your team stops paying for four tools that do not talk to each other, and the Monday standup runs off one board that was already true before the meeting.

What to try in the demo

  1. Tap an issue, cycle its status pill, watch the board agree
  2. Open Cmd K, type "done", execute; then type "billing" and jump to the issue it finds
  3. Filter the list to Urgent and find the two heavy conversations
george@navicade.comTell us what you are building. We answer within one business day.
← All the work

Payments console

Ledgerline

Built for: Any business moving money online and living with the fraud that follows it.

The problem

Payment operations live in the gap between "the chart says fine" and "this specific charge is fraud". The tool has to hold both: the aggregate view and the single suspicious payment with its evidence, plus the review queue where a human decides.

What we built

Books that reconcile. The volume tile, the chart and the payments table all sum from the same data, to the cent. Check it; reviewers do.

Risk with reasons. The risk score is computed live from the same signals shown on the card: CVC, address, IP against card country, amount. The queue names the exact signals that fired, and clearing a flag takes a reason that lands in the timeline.

The boring 85%. Most payments succeed unremarkably, because real books are boring, and boring is what credible looks like.

Decisions we made

Payment IDs are shaped exactly like the real API’s, and failure reasons are the ones that exist, because fintech buyers notice.

On Monday morning

The chargeback that used to eat a morning becomes a two-minute read: the timeline, the signals, the rule it tripped, and a decision button.

What to try in the demo

  1. Filter payments to Needs review
  2. Open one and copy its ID, then read its risk checks
  3. Approve a flagged payment with a reason, reload the page, find the decision still on the record
  4. Clear the whole review queue and meet the empty state
george@navicade.comTell us what you are building. We answer within one business day.
← All the work

Product analytics

Northbeam

Built for: Any product or marketing team drowning in events and starving for answers.

The problem

Dashboards answer yesterday’s questions. Analytics tools earn their keep when you can ask a new question and get a chart back in seconds: rebuild the funnel, split by plan, read the retention triangle. That interactivity is the product.

What we built

A funnel you can edit. Remove a step, add another, and the conversion math recomputes and re-animates, always decreasing downstream, because funnels that leak upward are lies.

The retention triangle. Eight cohorts decaying convexly to their plateau, every cell tappable for its exact number.

Segmentation that splits live. One line becomes three by plan, and the table under it matches the chart exactly.

Decisions we made

Series carry weekend dips and gentle growth because flat generated data is the fastest tell of a fake dashboard.

A 300 millisecond shimmer on re-query, because instant reads as fake in analytics.

On Monday morning

The "should we build annual billing" argument ends in four taps: funnel to upgrade, split by plan, read the drop-off, decide with numbers.

What to try in the demo

  1. Remove a funnel step and watch the math recompute
  2. Tap cells in the retention triangle
  3. Split weekly actives by plan, then flip the dashboard to 90 days
george@navicade.comTell us what you are building. We answer within one business day.
← All the work

Practice platform

Hartwell & Vane

Built for: Law firms and any regulated practice where the software must know the rules.

The problem

Legal software is a domain-knowledge test. Trust money must never mix with operating money, deadlines are court-fatal, time is billed in tenths, and estate intakes hide the classic failure: a beneficiary form that quietly outranks the will.

What we built

The whole practice on one screen. Matters with unbilled work, deadlines counting down in red, and trust funds in a dashed box that never sums with anything else. That separation is the law, drawn.

Time that becomes money. A running timer files entries in tenths at each person’s rate; the billable toggle moves totals instantly; one tap drafts an invoice that equals the time tab exactly.

The intake module. The estate questionnaire runs the conflict engine on the client’s own answers: add an insurance policy with no beneficiary and watch it get caught, or override a flag with a reason that goes on the record.

Decisions we made

The intake collects and organizes; the attorney does the legal work. No document generation, no advice, by design.

On Monday morning

The intake that took three weeks of chasing happens on the client’s couch, arrives pre-flagged, and every minute spent on it is already on the invoice draft.

What to try in the demo

  1. Start the timer, stop it, find the entry it filed
  2. Open the estate matter, run the intake, add an insurance policy with no beneficiary
  3. Generate the invoice and check it equals the time tab
george@navicade.comTell us what you are building. We answer within one business day.