AI platformsDistributed systemsResponsible AI

Challa Ravindranath. 20 years in the first seven layers. Now building for the eighth.

AI platformsDistributed systemsResponsible AI

  1. 8People
  2. 7Application
  3. 6Presentation
  4. 5Session
  5. 4Transport
  6. 3Network
  7. 2Data link
  8. 1Physical

Work

Four calls.

What I decided, and what it changed.

  1. 01 · Microsoft · Enterprise AI · 2025 to 2026

    Build the agent once. Let it run anywhere.

    I set platform and AI direction for a team of fifteen, and designed the agent harness.

    The call

    I exposed the skills headless, as an API, so the same agent plugs into the tools customers already use. The tool contracts are versioned, and the model only reaches data through them.

    The outcome

    One integration. We don’t rebuild it for each new surface.

    Every write validated. Nothing reaches the data without passing the tool contract.

    One request path Headless
    AssistantChatYour app
    Skills APIHeadless · serverless
    Agent harnessMemory · guardrails · models · tools
    Tool contractvalidated
    Enterprise dataSystem of record

    Same agent, same contract

  2. 02 · Microsoft · 2017 to 2025

    One ID across every boundary.

    One ID, end to end.

    The call

    I put one correlation ID across the front-end / back-end boundary where tickets got lost.

    The outcome

    Support SLA down 60%. And the whole organisation took it as its tracing reference.

  3. 03 · Microsoft · 2017 to 2025

    A text change isn’t a code change.

    Copy moved to the business; developers went back to code.

    The call

    Every text edit on our marketing sites was a pull request, a review and a deployment. I proposed moving content out of the codebase into a CMS, and drove it end to end, from the idea to a partnership with Contentful.

    The outcome

    Business users own their words. A copy change no longer needs a developer.

    Locales through content. A new market isn’t an engineering project any more.

  4. 04 · Roche Diagnostics · 2015 to 2017

    Correctness, on a live clinical line.

    Instruments processing about 1,400 samples an hour, where a flaky failure ships straight to a live diagnostic line.

    The call

    As system integration lead, I owned the merge from component release to the customer’s lab, including deploying and testing the full system on-site at labs across Germany.

    The outcome

    Stable releases across cobas 8100, p 612 / p 512 and p 501 / p 701.

Open source

Built in the open.

Contributor · Microsoft Agent Framework

What I build when nobody’s asking. Real AI systems, made in the open and measured on real data.

01 · Foundations
Underneath every AI system

AI Playbook

Data structures and system design, worked through the way real AI systems use them.

ai-systems-playbook · one request, end to endFull size ↗
One AI request flowing through a gateway, an agent, context and retrieval, LLM serving and streaming back, with the problems and algorithms at each part

Giovanni Paolo Panini, Ancient Rome, 1757 · The Met, public domain

Underneath every AI system are old ideas

10

families of classic algorithms, mapped onto the AI systems that run on them today.

Most people call an LLM API. I wanted to know what happens underneath.

Weeklya new problem, with a post
C# + testsevery solution runs
C# · .NET 10GitHub ↗
02 · Understanding code
What your code already knows

TSight

Point it at a codebase and see the riskiest file to touch, and the one person who knows it.

TSight · sample codebaseFull size ↗
TSight overview of a sample codebase: the riskiest file to change, knowledge held by one person, files nothing imports, and a map of change hotspots

Bernardo Bellotto, Pirna: The Obertor from the South, mid-1750s · The Met, public domain

What your code already knows

0

AI model calls. Every number is measured from the code and its git history.

Joining a team, you spend days building a map of the codebase in your head. Most of it is already in the code and in git.

0 bytesleave your machine
3 waysCLI · Docker · GitHub Action
TypeScript · ts-morph · gitGitHub ↗
03 · Retrieval
Where RAG pipelines waste money

Rechunk

Stops a company paying twice to teach its AI the same page. Only new or changed text gets embedded.

BenchmarkIn your pipelineFull size ↗

$ dotnet run -c Release --project benchmarks/Rechunk.RealData

markdown files on disk: 13555

files used: 2500 (sample 2500, seed 2026)

chunks in54,085

exact repeats · SHA-25611,610

near · SimHash21

near · MinHash44

skipped, never embedded11,675

left to embed42,410

most repeated · copies skipped

1,636## See also

567## Syntax

480## Requirements

396## Remarks

323## To correct this error

320## Parameters

245## Example

dotnet/docs @ 82937f4 · SimHash ≤ 3 bits · MinHash ≥ 0.85 · guards on

Where Rechunk sits in a RAG pipeline: after the chunker and before the embedder, sending new or changed chunks on, skipping duplicates, quarantining blocked PII, with deletes on a priority lane

John Frederick Kensett, Lake George, 1869 · The Met, public domain

Where RAG pipelines waste money

21.6%

of 54,085 real documentation chunks never reached the embedding model. They were copies.

Most pipelines re-embed every document on every ingest. You pay again for vectors you already have.

0 / 300number or “not” changes wrongly merged
99.5%precision at 92.6% recall, synthetic
C# · .NET 8v0.1GitHub ↗
04 · Answers from data
Where company data goes unasked

Pulse

Ask your company’s telemetry a question in plain English and get the answer with a chart.

Pulse · sample telemetryFull size ↗
Pulse answering plain-English questions about sample telemetry: a week of daily events, two orgs flagged below their 7-day baseline, summary cards, a 14-day trend and the ten quietest orgs

Canaletto, The Grand Canal, Venice, looking toward the Rialto, 1730s · The Met, public domain

Where company data goes unasked

85%

of plain-English questions are answered without an AI model, in under 50 ms.

Teams sit on huge amounts of telemetry, yet getting a simple answer is hard. Most of the time, people just don’t ask.

1.9 sanswer with chart, in the demo
30×faster follow-ups from cache
Python · Kusto · GPT-4oGitHub ↗

More on the bench.

New repos land here as they ship.
All repos on GitHub ↗

Beliefs

Trust takes time. Building it in from the first design saves it.

  1. 01

    Trust is architectural.

    You can’t bolt it on at the end. I put it in the first design, because that’s the only place it sticks.

  2. 02

    Correctness has a face.

    In a clinical lab, a failing test is a patient. I’ve never stopped seeing it that way.

  3. 03

    Prevent the category, not the instance.

    When the same bug came back, I stopped fixing the bug and fixed the design that let it in.

Recommendations

In their words.

Microsoft

“A thoughtful collaborator who uplifts those around him. His combination of technical depth, adaptability, and natural leadership make him an incredible asset to any team.”

Joel KatzevmanPrincipal EM, Microsoft · 5+ years together · Hyderabad, 2017 to 2026

Nokia

“Technical and business acumen always at the high end, a role model for his peers. A true professional and a true gentleman. Would work with him again without hesitation.”

Alexandru CicortasManaging Director, Emposo · ex-Nokia Helsinki · Helsinki, 2010 to 2014

Nokia → Microsoft

“Actively driving test automation across both protocol-level messaging and the user interface. Versatility and out-of-the-box thinking, an exceptional knack for identifying critical defects early.”

Haresh ChandaniHead of Cloud Engineering, KONE · ex-Nokia & Microsoft · Espoo, 2014 to 2015

Human in the loop

This is the step I never let an agent skip.

So this one is yours. Hold the button, and watch what trust looks like.

Press and hold, or press Enter.

waiting for your approval

For teams who refuse to ship what they can’t explain.

01 · Microsoft · Enterprise AI · 2025 to 2026

Building the agent harness

Contracts, not trust.

The problem

An AI agent on an enterprise platform has to touch real business data. One wrong write is an audit problem.

The call

I gave the agent one governed path, from conversation to model to data:

The value

02 · Microsoft · 2017 to 2025

Tracing that became the standard

Built so no one could get it wrong.

The problem

One business operation crossed back-end services, plugins, a data platform and a React front end, with no end-to-end trace. Every team had its own correlation pattern, and support tickets vanished at the front-end / back-end boundary.

The call

I put one correlation ID across that boundary:

The pattern, twice more

The value

60%

The support SLA was cut by 60%, and the whole organisation adopted the approach as its tracing reference.

03 · Microsoft · 2017 to 2025

Taking content out of the codebase

A text change isn’t a code change.

The problem

On the marketing websites, every text edit was a developer change. Copy that a product manager should own went through code, review and deployment. Marketing sites change often, so developer time was going into text, not code.

The call

I proposed the idea of a CMS and drove it end to end:

The value

0

Pull requests for a text change. Developer time went back to code, business users owned their own words, and the marketing sites could change at the speed marketing needs. The real call was seeing that this was never an engineering problem.

04 · Roche Diagnostics · 2015 to 2017

Clinical systems integration

From component release to the customer’s lab.

The problem

An instrument like the cobas p 512 is many components built by separate teams: the control unit, sample supply and module sample buffers. None of them counts until they work as one instrument on a live clinical line.

The call

As system integration lead, I owned the merge, from component release to the customer’s lab:

The value

~1,400

Instruments processing about 1,400 samples an hour, where a flaky failure ships straight to a live diagnostic line. Integration was what made separate components safe to run in a real lab.