Shraddha Yeolekar
AI agent deployment · Customer experience · Regulated industries

I launch AI that customers can trust.

For five years, I have put AI chatbots in front of real customers at a regulated trading company. I test the AI before launch, set a bar it must clear, and say "not yet" when it is not ready. Below are three of my projects, with the key decision in each and how it works.

Shraddha Yeolekar
Where I've done this for real

At OANDA, a regulated online trading company, I owned the AI and automation portfolio for client experience: chatbots, autonomous agents, and the human review that keeps them accountable.

+25%
Monthly case deflection from the first client-facing AI chatbot
<1%
Hallucination rate held since that launch
0
PII or unauthorized-advice incidents to date
30+
Conversation flows designed, including the hand-off to human agents
Project 01 · Document AI

AI Document Pipeline

Key decision

Don't ship it yet.

The tuned pipeline scored 82% against an 85% bar, with contracts at 60%. The eval gate held it back.

EvaluationRAG
Uploaddocument Longdocument? no: whole document Gemini Flashclassify + extract confidence≥ bar? auto-approve human review long docs yes Section-awarechunks Query per fieldtop 3 each per-field context every change Configchange Eval gategolden set pass Live configprompt · chunks · K sets fail → stays pending
How a document moves through the pipeline (top two lanes) and how a change reaches production (bottom lane). Nothing in the live configuration changes without passing the gate.
Project 02 · Live public AI

Ask AI

Key decision

Run the cheapest, most predictable guardrail before the model.

Risky questions are refused by a keyword filter before Gemini is called, and before they count against a visitor's daily limit.

GuardrailsLive
before the model · deterministic after the model Killswitch Spamcheck Keywordpre-filter Rate limit5 / day GeminiJSON + cites Citationcheck Answer off → paused bot → rejected risky → refusal no model call, no quota spent 6th ask → tomorrow unknown link → dropped refusals and answers are logged to a database for review
Every check before the model can stop a request. The highlighted pre-filter is the key decision: it costs nothing, behaves the same way every time, and runs before the rate limit.
Project 03 · Agentic AI

Multi-Agent Support Orchestrator

Key decision

Upset customers and exception requests go to a person, not the bot.

That's where a fluent wrong answer does the most damage, so the bot writes a hand-off summary instead of a reply.

AgenticHuman-in-the-loop
Customermessage RouterKB or ESCALATE policy question Vector searchbest-matching fact KB agentanswers from it → reply to customer angry or exception Escalation agentsummary + urgency Human agentpicks up ticket every model call is traced: latency, tokens in and out
The highlighted path is the decision: the riskiest conversations skip the bot's answer and reach a person with a summary.

6 projects in total, each with its decision and design.

See all projects →
Live · ask.shraddhayeolekar.com

Don't take my word for it. Try to break it.

Ask AI answers questions about my work using only my resume and GitHub projects, and cites its sources. These are the kinds of questions it's built to refuse. Try them, or try your own.

Open Ask AI →