Avataar · 2024 to 2025 · Proof of concept

The handoff is a form, not an error.

A 0-to-1 agentic platform for regulated work. When it isn't sure, a person doesn't get a failure. They get a decision already prepared.

RoleAI Product Manager
DatesJul 2024 to Oct 2025
DomainHealthcare, medical encoding
StageFour workflows, POC
Escalated for review62%
From the record
"...persistent atrial fibrillation, rate controlled, on anticoagulation..."
Suggested
I48.21
The rule in question
Chronic vs persistent distinction changes the sub-code.
Alternatives
I48.20I48.19I48.11
AcceptChoose anotherSend back

Illustrative. What it never says is "task failed".

80%+Task accuracy on the medical encoding POC
80 / 20Handled autonomously, escalated with context
<4sP95 per agent step, end to end
3Deployment modes from one codebase

The problem

An agent that improvises can't be allowed near a medical record.

The appeal of agents is that they work things out as they go. In a regulated workflow that's exactly the problem. You can't approve a plan that doesn't exist yet. You can't price work you can't see. You can't contain a step nobody decided on.

And when it goes wrong, "the agent decided to" isn't an answer anyone in healthcare accepts.

Everything here follows from refusing to let the system make it up as it goes.


The first decision

The plan is written before anyone starts working.

The orchestrator doesn't improvise step by step. It produces a plan up front, showing which steps depend on which. The cost is flexibility. What you buy is worth more.

read record
extract diagnoses
find procedures
look up codes
validate
ReadA plan a person can look at before it runs.
PriceKnown steps means known cost, in advance.
ContainLimits apply to steps that already exist.
StopBefore anything happens, not after.
Point atWhen it goes wrong, a specific step, not a vibe.

The shape of it

A small team with strictly defined jobs.

Five layers, and the value is in what each one is not allowed to do. Only one of them can touch an outside system.

Orchestrator
The manager

Receives the job and writes the plan before anyone starts working.

Task agents
The specialists

One extracts the diagnoses. One finds the procedures. One looks up codes.

Tool agents
The hands

The only ones that touch outside systems. Everything else reasons.

Validator
The checker

Confirms the output is well-formed and permitted.

Sentinel
The safety officer

Watches for things going wrong, and can stop the run.

Memory is two tiers, deliberately separated. Short-term holds the current job and is thrown away afterwards. Long-term persists what is worth keeping. Keeping them apart is what stops one case bleeding into another.


Governance

Three mechanisms, not a policy document.

Each of these is something the code enforces. None of them is something a document requests.

Contain

Blast radius

Limits on how much damage a single action can do. The bound exists before the action does.

Diagnose

Step-level traces

A record of every step, so a wrong action is explicable afterwards rather than mysterious.

Prevent

Confidence floors

65% while testing. 85% minimum in production. Retuned against what actually happened, not fixed at launch.

The decision I would defend hardest

Review people use, rather than review people route around.

Below the confidence floor, the work goes to a person. What that person receives is the whole design.

Not "task failed". A pre-filled decision form, with the context already compiled. The relevant record excerpt. The suggested code. The rule in question. The alternatives.

That single choice is the difference between review people use and review people route around.

It's the same idea that shows up later in document processing. Escalation is a product surface, not an error path. Build it as a failure and people will work around it. Build it as a decision and they will use it.


Where it stopped

The proofs of concept never converted.

01
Four workflows, fully automated, and none of them became a live client.

They were proofs of concept in medical encoding, and they had not converted before I left. The accuracy figure is sound. Its scope is a POC. Worth stating plainly, rather than letting a reader assume otherwise.

02
The question underneath is whether it was too generic to sell.

A horizontal agent platform sold into medical coding competes with vendors who ship the vertical, and that tension is real, and having a view on it is more useful than avoiding it.

03
Screen-based agents for systems without APIs are brittle by construction.

Where the work lived in legacy systems with no integration, agents reasoned over the screen. It unblocked the work and a layout change is an outage. Nothing records where that actually bit.

04
The ceiling on the 80% autonomous share was never established.

Eighty per cent handled without a person. That's the observed split, not a limit anyone tested for, and what it would take to move it, or what it would cost, isn't on the record.

What happened to it

The architecture outlived the context it was built in.

Three deployment modes came out of one codebase: cloud managed, hybrid, and air-gapped on premises. That last one is what regulated buyers actually ask for. It's a decision you make at the beginning or not at all.

Foundation models were routed per workflow by accuracy and cost rather than picked once: GPT, Claude and domain-specific models, with the platform agnostic to all of them.

The same architecture is the foundation of a different company's platform now. That says more about the design than any accuracy figure does.

Product, architecture and governance design: mine. Built with a cross-functional pod at Avataar. The reference architecture described here is generic; implementation figures are POC-stage and scoped as stated.

Working on something that isn't obvious yet?
Let's build it.

Shubham Shrivastava
Shubham ShrivastavaHead of Product · Neoflo.ai
© Shubham Shrivastava 2026