Eupraxic

Articles

The dashboard panel a hospital CMO trusted most was the one showing failures

Telemetry from the first commit, a feedback loop that runs at three speeds, and why the honest panel matters more than the metrics.

By Nasser Ghanemzadeh · · 5 min read

In week ten of a clinical-documentation deployment I've seen at a regional health system, the CMO told the team which dashboard panel she valued most. She passed over acceptance rates and turnaround times. She picked the panel that named one thing the system had gotten wrong, and what the team was doing about it.

That panel exists only if the loop behind it runs. In classical software, testing happens before deployment and production monitoring catches what the tests missed. AI systems don't allow that split. Charity Majors made the case at Honeycomb in May 2023: nondeterministic outputs collapse the line between testing and observing, so you ship fast, instrument everything, and feed production traces back into prompts and evals.

In a forward deployed engagement that loop runs at three speeds at once: per trace, per week, and per month.

Instrument from the first commit

Telemetry is what the system is for, starting with the first slice. Every model call, tool invocation, user input, output, human intervention, and escalation gets a structured log line with enough context to rebuild the request later. Later often means six months on, after two model swaps and a hundred prompt edits, when someone asks why the system made a particular recommendation on March 14.

The minimum schema is unexotic: a trace ID, the agent's identity and version, the input with whatever PII handling governance requires, the pinned model version, a prompt hash, each tool call with its arguments, result, and latency, the output, any eval score, where a human stepped in, and the token cost. What matters is producing it on every single trace.

Send it to the customer's observability stack, whether that's Datadog, Honeycomb, Splunk, or an LLM-specific tool like Langfuse. The customer's SRE team will read these traces, so they have to live where that team already works.

The per-trace loop, and the regression it caught

The per-trace loop is daily work. Sample production traces, weighted toward what the inline judge or the customer flagged. Read them until a pattern shows. Turn the pattern into an eval case, code the fix, and ship it only if the eval suite says it's better. Hamel Husain's insistence on reading actual traces yourself, instead of staring at aggregates, is the heart of it. (Evals got their own piece earlier in this series.)

Here is what that looked like six weeks into production at the same health system. The dashboard showed first-pass acceptance (notes signed without any edits) drifting down. It had settled at sixty-one percent in week three. By week six it was fifty-four. On a noisy weekly chart, that slope was easy to miss.

Two sessions found the cause. In the first, the team pulled forty traces the inline judge had flagged as low-confidence. Thirty-three came from the orthopedic clinic, which had only come online in week four. Orthopedic notes run shorter on history and longer on physical-exam findings, with a structured assessment the prompt had never seen. The drafts were accurate, but the orthopedists were rewriting them. What looked systemic was one specialty.

The second session was a labeling pass with the lead orthopedist: forty cases, two hours. Her asks were concrete. Order the exam section by joint rather than by finding type. Use the primary diagnosis, differential, and plan structure she had learned in residency. Cut the history section way down. The prompt revision took an afternoon and shipped on a Friday. The next week, acceptance was back at sixty-three percent, with orthopedics tracking family practice for the first time.

One more detail matters. The judge had been calibrated on family-practice notes. Its low-confidence flag still pulled the right traces, but its absolute scores on orthopedic notes couldn't be trusted, so the team recalibrated it on a fresh orthopedic sample in the same cycle. A rubric is never finished. Every new population with its own definition of good has to be encoded, and the per-trace loop is what tells you when.

Memory lives in the wrapper

The per-month loop carries field learning back into the product. The engineer writes a two-page synthesis of changes made, eval cases added, and open questions the product can't answer yet, and the product team picks one to three items to build into the platform. Daniel P.'s synthesis of FDEs at Palantir, Distyl, and Modal calls the productive tension between those two groups managed misalignment.

Inside the engagement there's a quieter problem. Andrej Karpathy, in his Software 3.0 talk, called it anterograde amnesia: agents don't learn from one session to the next. The model doesn't remember that compliance rejected a phrasing two weeks ago, or the corrections the customer's expert made to the eval set. Whatever memory the system has lives in the wrapper. The wrapper encodes accepted corrections, carries state within and across sessions, and loads the customer-specific context the model can't otherwise reach.

That makes the forward deployed engineer the institutional memory of the engagement. The customer's team will rotate and the model will be replaced. A well-documented wrapper survives both. At handoff, the package should include a runbook, the eval suite with its rubric history, the prompt store with the rationale for each major change, dashboard documentation, an escalation playbook, and a calibration session with the customer's team. A clean package keeps the system performing. A messy one lets the wrapper, the evals, and the context drift apart within months. Write the wrapper for the engineer who will replace you.

The dashboard that built trust

The per-week loop is the one the customer sees, and its artifact is the dashboard. Build it on the success criteria from scoping, in the customer's units: tickets resolved, minutes saved per claim, calls handled per dispatcher. Walk the sponsor through it for fifteen minutes every week for the first six weeks of production. The metrics matter less than the conversation they make possible.

The CMO's dashboard had five panels: notes drafted by service line, first-pass acceptance with confidence bands, median minutes from visit end to signed note, the week's top three edit reasons in plain English, and one named failure mode with the planned fix, written by the engineer.

The first real failure shown there was an ugly one. The system had turned "no known drug allergies" into "known drug allergies" in three notes before the prompt was fixed. Clinicians caught and corrected all three before signing. The dashboard named the failure, the prompt change that closed it, and the new eval case that guards against that class of error.

In the meeting, the CMO said she had stopped reading vendor status reports years ago because vendors hid the bad news. This dashboard was the first artifact she trusted. If the failure rate on a dashboard doesn't match what users experience, the dashboard is worthless. When it does match, you can ship the next iteration without re-arguing the whole deployment.

The metrics are the medium. The honesty is the message.

Thinking about moving into this work?

I wrote a free guide for people considering forward deployed engineering: what the 4.6x growth figure means, how to tell a real FDE role from a renamed sales job, and a 30-day plan that ends with your first piece of proof.

You get the PDF by email, then occasional notes on making the switch. Unsubscribe in one click. We use your details only to send these emails. Privacy policy

See what is inside

About the author

Nasser Ghanemzadeh is a former CEO and CPO with 17 years of product leadership. He exited Nivo, a B2B fintech with 200K+ users, after growing revenue 450%, was founding head of product at Pangouan, and founded the Finnova accelerator, which he ran as its first managing director. He builds Vectig with Claude Code and MCP every day and runs every Eupraxic Sprint himself. Eupraxic is one engineer for now; the method is written down so it will not stay that way.

This article draws on his book, Forward Deployed AI Engineering.

Want one useful AI workflow in your product?

Score your workflow in two minutes, then book a 15-minute Scoping Call. If it fits, you leave the call with a Brief draft and a kickoff date.