AI Engineer - Harness Engineering - Agent Observability - Founder
ScrollAbout
Most AI failures aren't model failures - they're orchestration failures. I build the harness around the model: the decision layer of routing, retrieval, tools, memory, guardrails, and evaluation that makes AI reliable in production.
Nearly three years shipping production LLM, RAG, and agentic systems for enterprise customers - the work that decides whether AI holds up: retrieval quality, evaluation, guardrails, and latency.
Harness engineering
Most AI failures aren't model failures - they're orchestration failures. The hard part of production AI isn't choosing an LLM; it's deciding how every request flows through models, tools, data, memory, and guardrails to produce a reliable outcome. That decision layer is the harness.
The LLM generates the answer. The harness makes sure the right answer is generated in the right way.
A production harness is the decision layer that determines:
How I think about production AI - a reference model, not a single product. Every request flows down through the layers; observability and evaluation wrap all of them.
Authenticate, classify, route, and screen every request before a model ever runs.
Pull the right knowledge from the right source before generation.
Decide which tools and agents act, and in what order.
Validate, attribute, and score before anything reaches the user.
Not just a chatbot - an AI operating system that routes every request through the most reliable execution path.
The future of production AI won't be defined by the largest model. It will be defined by the quality of the harness built around it.
Experience
Five roles, one thread: shipping AI systems - and the harness that keeps them reliable.
Microsoft
Apr 2026 - Jun 2026
Full-time
Built tool-calling agents and harness engineering for the Azure Compute org - the decision layer that turns a model into a reliable, unattended workflow.
Aug 2024 - Apr 2026
Full-time
Forward-deployed engineer for Fortune 50 finance and healthcare. Stood up the harness observability and evaluation that lets teams see - and trust - what their AI is doing.
QSS Technosoft
Aug 2023 - Aug 2024
Full-time
Shipped production conversational AI for US clients - prototype to customer production.
Celebal Technologies
Jun 2023 - Aug 2023
Automated the data and ML plumbing that feeds the models - less manual prep, faster training loops.
Mar 2022 - Aug 2024
Founded an AI mental-health platform with real-time voice and an LLM backend, and made the product and architecture calls end to end.
What I build
The work spans the harness - from agents and retrieval to the observability and guardrails that keep them honest.
The decision layer around the model - routing, retrieval, tools, memory, guardrails, and evaluation. Built tool-calling agent harnesses at Microsoft.
You can't trust what you can't see. Stood up tracing, monitoring, and LLM-as-a-judge evals at Fiddler so regressions surface before users feel them.
PII detection, policy enforcement, and defense against prompt injection and hallucination. Built guardrail and red-team frameworks for Fortune 50 AI.
Hybrid (vector + BM25) and graph RAG, tuned where generic pipelines fail on domain data. Retrieval is what decides whether the answer is grounded.
Route each request to the right model and workflow by quality, latency, and cost; orchestrate tools and agents into reliable multi-step flows.
Self-hosted vLLM for predictable tail latency, plus the CI/CD and plumbing that keeps AI systems up. The unglamorous work that matters.
The toolkit
The stack behind the harness - languages, agent frameworks, retrieval, and the observability that keeps it honest in production.
The arc of the work
No invented proficiency scores - the signature domain each chapter centered on, in the order it happened. Five roles, one direction - now open to the next.
Writing & research
Two papers presented at IEEE / IIT Delhi and Poornima University - plus certifications, kept current.
Speaking
From presenting papers at IIT Delhi to returning as an alumnus to the room where it started.
“Where you start doesn’t decide where you end up.”
Alumni talk - where I started · Manav Rachna University
Went back to the same room where I sat five years ago. Told students what I learned. How to get curious about hard problems. How to work with people who push you. How your starting point does not define your end point.
The journey
Built in parallel - a startup carried alongside study and full-time roles.
Contact
Open to conversations about agentic AI, enterprise AI systems, and the occasional good idea worth chasing. Drop a line - it’s one of my favourite things to talk about.
Prefer email? letstalk@heytaha.com LinkedIn Resume
