AI engineer, Helsinki 60.17° N, 24.94° E
Manisha Parwani.I build agentsyou can trace.
Claude and Azure AI systems that go to production with evals, guardrails and tracing, so the people who rely on them can see why a decision was made.
What the job is now
The agent isthe easy part.Knowing why itdecided is the work.
I spent eight years owning enterprise software from design to production: releases, on-call, root-cause analysis. AI did not change that job. It made it the whole job.
Fourteen years, one trace.
Each bar is a span on a real time axis. Select one to read its attributes.
- 4.0y
- 4.4y
- 2.8y
- 3.1y
- 0.3y
- 0.5y
- 1.7y
- 1.4y
- 1.0y
Spans worth opening
All work- 2025A second opinion for every AI decisionSanoma Media Finlandagent.second_opinion
- 2025From booking to live, without the spreadsheetSanoma Media Finlandsoftware.engineer
- 2021Keeping a global corporate site steadyAccenture for Avanadesenior.software.analyst
- 2026An agent that reads the logs firstLab buildlab.root_cause
- 2026Runbooks that answer back, with a scorecardLab buildlab.runbook_rag
What I bring to an AI team
Agents and evals on top, the platform they run on underneath, and the habits that keep both honest in production.
Agents and LLM apps
Tool-using agents that do real work inside a business system.
- Claude Agent SDK
- Claude Code and Skills
- Model Context Protocol servers
- Tool use and structured outputs
- Microsoft Agent Framework
- Semantic Kernel
- Azure OpenAI
- Microsoft Foundry
- Multi-agent orchestration
- LangGraph
- Prompt engineering
Retrieval and documents
Getting the right context in front of a model, and proving it was right.
- RAG pipelines
- Embeddings
- Azure AI Search
- pgvector
- Chunking and reranking
- Azure AI Content Understanding
- Document intelligence
Evals and safety
Nothing ships on vibes.
- Golden datasets
- LLM-as-judge
- promptfoo
- Human-in-the-loop review
- Guardrails
- Azure AI Content Safety
- Prompt-injection defence
LLMOps and observability
Every model call is a span with a cost, a latency and a reason.
- OpenTelemetry GenAI tracing
- Langfuse
- Application Insights
- Log Analytics and KQL
- Cost and latency budgets
- Model routing
- Evals in CI
Platform
The cloud the agents run on, owned end to end.
- C# and .NET
- .NET Aspire
- TypeScript and Vue
- Python
- Azure Functions and Durable Functions
- Container Apps
- Service Bus
- API Management
- Front Door
- Bicep
- GitHub Actions
- Azure DevOps
- Entra ID, RBAC, OIDC
- SQL Server, MySQL, Oracle
How I work
The part that does not change when the model does.
- Root-cause analysis
- Production support
- Release management
- Architecture decision records
- Requirements analysis
- Scrum
- Team leadership
- Vendor coordination
How I ship
AI systems.
Trace every decision
A model call is a span: input, output, cost, latency and the reason. If I cannot replay why an agent did something, it is not ready.
Give the model a second opinion
Agents propose, people decide, and the disagreements are the most useful data in the system.
Evals before opinions
A golden set and a judge in CI settle arguments about prompts faster than any meeting.
Own it to production
Infrastructure as code, least-privilege identity, a release checklist and a runbook. The demo is day one.
Questions recruiters ask
Who is Manisha Parwani?
Manisha Parwani is an AI engineer based in Helsinki, Finland. She works at Sanoma Media Finland, where she builds AI validation of advertising creatives with Azure AI and a Claude Agent SDK agent that gives human reviewers a second opinion. Before that she spent eight years delivering enterprise .NET and Azure systems in Mumbai, including leading a Sitecore team at Accenture.
What does Manisha build?
Agents you can trace: LLM applications and tool-using agents on Claude and Azure AI that ship with evals, guardrails and OpenTelemetry tracing, deployed with infrastructure as code and least-privilege identity.
Which AI tools and platforms does she work with?
Claude Agent SDK, Claude Code, Model Context Protocol, Azure OpenAI, Microsoft Foundry, Azure AI Content Understanding, Azure AI Search, Semantic Kernel and Microsoft Agent Framework, with promptfoo and Langfuse for evals and tracing. The platform underneath is .NET, TypeScript, Python, Azure Functions, Container Apps, Bicep and GitHub Actions.
Is Manisha open to new roles?
Yes. She is open to AI engineering roles in Helsinki and remote across the EU. She holds an EU Blue Card and lives in Finland.
How do I contact Manisha?
Email hi@manisha.dev or use the contact form on manisha.dev. She also answers on LinkedIn.
Open a span
Tell me what
you are tracing.
Open to AI engineering roles in Helsinki and remote across the EU. I read every message myself and reply within two working days.
hi@manisha.devOr on LinkedIn