Barjinder Singh Agentic AI Engineer· Open to full-time roles
I build agents, and evals that break them.
Python, LangChain, LangGraph, MCP and Azure OpenAI. Before that, seven years of QA on ML systems at Wipro, Amazon and LTIMindtree. Go on, move your cursor.Go on, tap him.
I write agents in Python that call tools, keep context across turns and
pass work to other agents when that is the right move. Two of them are
on this site: a RAG pipeline with the cosine-similarity search done by
hand, and a supervisor that routes questions to three specialists. Both
carry an evaluation harness and a JSONL trace log with the cost and
latency of every call. I wrote my first agent loop by hand before
picking up LangChain, so I know what the framework does underneath.
Before that I spent seven years on quality: validating datasets at
Wipro, auditing ML-integrated data systems at Amazon, then evaluating
model output at LTIMindtree. It transfers. An agent that works on Monday
and fails on Thursday is a flaky test with a bigger bill, and
reproducing the failure before touching the code is the same habit
either way. Oracle and IBM have certified me in agentic AI. The seven
years are what I would point at.
Skills
Skills
What I work with
Agentic AI & LLMs
Autonomous agents
Tool / function calling
Agent loops
Multi-agent orchestration (supervisor pattern)
Prompt engineering
LLM evaluation
Frameworks & protocols
LangChain (create_agent, @tool)
LangGraph (checkpointer, thread_id)
langgraph-supervisor
OpenAI Agents SDK
Model Context Protocol (MCP)
RAG & retrieval
RAG pipelines
Embeddings
Cosine-similarity search, from scratch
Document chunking
Grounded generation & refusal
Embedding caching
Evaluation & reliability
Evaluation harnesses
Regression testing
Cost & token tracking
Latency tracing
Root cause analysis
Prompt-injection testing
Providers & tooling
OpenAI
Azure OpenAI (keyless / DefaultAzureCredential)
Google Gemini
Oracle Cloud
REST APIs · JSON / JSONL
Git · Linux / WSL
Languages & data
Python
SQL
JavaScript
Java
PHP
DBMS
Web & interfaces
HTML
CSS
Bootstrap
React
DOM
Ways of working
Jira
Confluence
Slack
Experience
Experience
Where I've worked
LTIMindtree
Senior Cloud Engineer, AI quality evaluationMay 2025 – Jan 2026
Evaluated AI and ML model outputs for accuracy, consistency and
reliability using Python and structured evaluation criteria.
Performed root-cause analysis on model failures and edge cases,
and authored the evaluation guidelines that standardised how
quality was measured across the team.
Amazon
Quality Auditor, QA on ML-integrated systemsJan 2020 – May 2025 · Hyderabad
Designed and applied structured evaluation criteria to audit
large-scale ML-integrated data outputs, catching errors before
they reached production.
Built repeatable QA checklists and acceptance criteria that
put "correct vs. incorrect" in writing, so every auditor drew
the line in the same place.
Ran systematic sampling across high-volume datasets to
estimate quality at scale without reviewing every record.
Treated recurring failures like flaky tests: reproduced the
failure, isolated what triggered it, then fixed that.
Categorised recurring defects into failure types, then
prioritised fixes by frequency and impact.
Drove a 25% reduction in recurring defects through root-cause
analysis, and built regex-based validation checks that caught
errors before delivery.
Sustained 96%+ audit accuracy while meeting throughput targets
in a high-volume production environment.
Wipro
Senior Data Associate, quality & data operationsJun 2018 – Dec 2019 · Hyderabad
Performed QA and validation on high-volume datasets feeding
downstream ML and analytics systems, applying detailed
guidelines and quality checks.
Certifications
Credentials
Certified, and verifiable
Make Agentic AI Work for You
IBM SkillsBuildIssued Jul 2026
Where agentic systems pay off and where they do not: task
decomposition, tool calling and putting a human in the loop
on the steps that need one.
Agent design, tool use and orchestration on Oracle Cloud
Infrastructure. The foundations exam for building and
running AI agents rather than only prompting a model.
Seven courses on putting AI into the work itself: research,
writing, content, data analysis and app building. Assessed on
a portfolio of artefacts and one AI tool built end to end.
Answers questions from what a document actually says. Chunking,
embeddings and cosine similarity all written by hand, with no RAG
framework underneath. When the document has no answer, it says so.
Eval trace 3 of 20 rows
{"q":"Which cloud did he use at Amazon?","hit@1":true, "hit@3":true, "top":"exp/amazon#3","cos":0.71}
{"q":"How many years of QA?", "hit@1":false,"hit@3":true, "top":"skills#1", "cos":0.44}
{"q":"What is his notice period?", "hit@1":null, "hit@3":null, "answer":"not in the document"}
Row 2 is why top-1 sits at 50% and top-3 at 100%. Row 3 is the
abstention path: no supporting chunk clears the floor, so it declines
instead of inventing a period.
A supervisor hands each question to one of three specialists: dates,
arithmetic, resume lookup. The eval scores which specialist got it,
so a right answer from the wrong agent still fails.
Routing assertions 3 of 5 cases
{"case":1,"q":"Days between 2024-01-05 and today?","routed":"dates", "expected":"dates", "pass":true}
{"case":2,"q":"12.5% of 48,000?", "routed":"math", "expected":"math", "pass":true}
{"case":3,"q":"Which employer used LangGraph?", "routed":"resume","expected":"resume","pass":true}
assert row["routed"] == row["expected"] # answer correctness alone never passes a case
Scoring the answer alone would have passed a run where the maths
agent guessed a date correctly. Asserting the route is what caught it.
Agents that plan, call tools and check their own output. Built
against a 30-case eval set, so "it works" is a number rather than a
demo that happened to go well.
Plain HTML, CSS and JavaScript. No dependencies, no build step,
nothing to patch at 3am. This site is one of them, including
the CV you can print from it.
For when the system has to remember who is using it. Accounts,
stored data, an admin view and a typed API, with the schema
and the API contract settled before any application code.
A pipeline, not a prompt box. Every ratio the placement needs,
with your logo and product composited in afterwards, so a model
never has to redraw them.
