Challa Ravindranath

Engineer at heart · Strategist at mind

Trust, by design.

Tactics are only worth the strategy they serve.

Focus areas

  1. 01

    AI systems

    Agents that act on real data, safely.

  2. 02

    Distributed systems

    Tracing and correlation that make failure visible across every boundary.

  3. 03

    Responsible AI

    Guardrails in the design, and every AI deliverable through review.

Beliefs

  1. 01

    Trust is architectural.

    You can't add it later. It is in the design, or it isn't there.

  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.

    If the same bug comes back, the design is wrong, not the code.

The call.
The outcome.

Every engineering decision serves a strategy. Here is the tactical call, and what it changed.

01 · Microsoft · Enterprise AI · 2025–26

Build the agent once. Let it run anywhere.

Set platform and AI direction for a team of fifteen, and architected the agent harness.

The call

Headless skills exposed as an API, so the same agent plugs into the tools customers already use. Versioned tool contracts: the model calls tools and never writes unvalidated data.

The outcome

One integration, reusable everywhere.No rebuild for each new surface.

Governed by design.Every write validated, and every AI deliverable through Responsible AI review.

  1. 02 · Microsoft2017–25

    One ID across every boundary.

    Support SLA cut 60%. Became the org-wide tracing standard.

    The call

    Correlation IDs across a front-end / back-end boundary that swallowed tickets.

  2. 03 · Roche Diagnostics2015–17

    Correctness, on a live clinical line.

    Releases held together on live clinical lines: four product families, ~1,400 samples an hour.

    The call

    As system integration lead, owned the merge of every component into one cobas instrument, and the validation framework behind it.

  3. 04 · Microsoft (ex-Nokia)2014–15

    If trust isn't in the handshake, nothing moves.

    Carrier certification with Orange, Deutsche Telekom and Vodafone.

    The call

    Protocol-level RCSe validation: SIP signalling against GSMA RCS specifications.

Built in the open

Contributor · Microsoft Agent Framework

Active C# · .NET · DSA · System design

Playbook

Data structures, algorithms and system design, worked through the way real AI systems use them: agent DAGs, KV caches, batching, RAG. Every problem comes with sources and tested C# code.

GitHub ↗
  1. An agent plan is a DAG.
  2. A prompt cache is a radix tree with LRU eviction.
  3. Packing a context window is a knapsack problem.
TypeScript · AST · git history

TSight

No model, no guessing. Every number checks against the repo.

One local tool, six lenses on any TypeScript codebase. Run on several large production codebases, and your code never leaves the machine.

Mental model
Architecture layers at a glance
Hotspots
Complexity × dependents
Blast radius
What breaks if a file changes
Bus factor
Who knows which code
Issues
Duplicates · cycles · unused files
Codebase map
Onboarding docs, exported
GitHub ↗
Azure Data Explorer · YAML · LLM

Pulse

The model asks. Config decides.

Plain-English questions to Azure Data Explorer. The model writes only the filter clause; clusters, schemas, auth and metric definitions live in config it can't touch.

GitHub ↗

In their words.

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.
Principal EM, Microsoft · 5+ years togetherHyderabad · 2017–26

I build teams that deliver value to the business, not just software. Looking for the next one.

Let's talk.

02 · Microsoft · 2017–25

Tracing that became the standard

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

One correlation ID across that boundary. It lives in the instrumentation layer, not in product code, is safe across threads, is enforced through the type system, and is wired into structured logging. The same instinct produced a design-token system that cut accessibility regressions, and a CI failure classifier that finally separated infrastructure flakes from real regressions. Different domains, one move: eliminate the category.

The value

60%

Support SLA cut by 60%. The approach became the reference for tracing across the whole organisation.
01 · Microsoft · Enterprise AI · 2025–26

An agent harness, governed by design

The problem

An AI agent over an enterprise platform has to touch real business data, where one wrong write becomes an audit problem.

The call

One governed path from conversation to model to data:

A memory layer holds conversation state.

A gateway applies guardrails and rate limits on every call.

Models run on a managed, serverless AI platform.

Enterprise data is reachable only through tools.

Skills are exposed headless as an API, so the same agent can plug into the tools customers already use.

Versioned tool contracts mean the model can call tools but never write unvalidated data.

Every AI deliverable clears Responsible AI review.

The value

An agent built once and reusable wherever customers work, with no second integration. Every write is validated, and every AI deliverable clears Responsible AI review. Direction set for a team of fifteen engineers.
03 · Roche Diagnostics · 2015–17

Clinical systems integration

The problem

An instrument like the cobas p 512 is many components built by separate teams: the control unit, sample supply, module sample buffers. They only count once they work as one instrument on a live clinical line.

The call

As system integration lead, own the merge:

Brought every component together into one integrated system build: the control unit (system coordination, LIS connectivity), sample supply and module sample buffers.

Did it across four product families: cobas 8100, p 612 / p 512 and p 501 / p 701.

Owned component releases through to the integrated build.

Led the cross-product Python validation framework, including the Python 2 → 3 migration, so it outlived the interpreter.

Later moved to the firmware team, working below the application layer on the instruments themselves.

The value

1,400

Releases held together on live clinical lines across four product families, running around 1,400 samples an hour.
04 · Microsoft (ex-Nokia Devices) · 2014–15

Carrier-grade protocol validation

The problem

Rich Communication Services on Windows Phone could only reach people once each carrier certified it, and certification is decided at protocol level: SIP signalling against the GSMA RCS specifications.

The call

Validate where the trust is decided, in the handshake:

Owned RCSe validation at protocol level, against GSMA RCS conformance.

Tests in C#, running in CI.

Logs and SIP traces as the diagnostic surface, so a failure is read in the messages themselves.

The value

3

Carrier certification with Orange, Deutsche Telekom and Vodafone.