Jannet Akanksha Ekka
Open to applied AI & AI evaluation roles · remote

I architect agentic AI
and keep it running.honest.measured.running.

I ship production AI systems by directing AI coding agents, and I am the only review before anything merges. Four years leading QA automation at Deloitte taught me how to tell when the output is wrong, which is the part most AI teams are still learning. I run a patent-pending AI trading agent, live 24/7 on a free-tier cloud stack. Rank 1 in my AI/ML postgrad at UT Austin McCombs.

See the work Get in touch

Résumés (PDF): Applied AI & AI evaluation · Founder — JTech Solutions / SMT

⚙️ Patent pending 🤖 AI-native delivery 🛡️ Agent security · Wiretrap 🏆 OpenServ Best DeFi — Winner 🥇 Rank 1 — PGP AI/ML ☁️ Google Cloud Gen AI Academy
25
Projects shipped
91K
Lines shipped & audited
4
Years enterprise eng.
1
Patent filed
Selected work

Systems, not notebooks.

Everything below either runs somewhere or ships with the tests that prove it works. Filter by discipline.

Flagship Patent pending · IN 202631090789 Seeking grants & partners

Smart Money Trading (SMT)

Explainable AI Trading Agent

Jan 2024 – Present · Founder & sole engineer · JTech Solutions (pre-incorporation)

An AI trading agent for 8 crypto perpetual-futures pairs that has to prove why it traded. Written rules with replayed records open every trade. Six AI advisors explain each decision and are graded on whether they were right; one earns a veto only after 30 proven warnings. Before a reason reaches a user, the system flips that advisor’s input and re-runs the decision, and drops the reason if nothing changes. The learning loop retunes only through a CPCV + Deflated-Sharpe + PBO + FDR overfitting gate. It places real orders 24/7 on WEEX’s exchange-granted simulated account from a free-tier Oracle Cloud VM, and SMT World, a Three.js/WebGL ocean on Cloudflare Workers, lets anyone watch it trade and ask it why.

91Klines of Python
3,045tests green
611findings recorded
344PRs merged
Multi-Agent SystemsExplainable AIAI EvaluationAI-Native DevelopmentOracle Cloud + systemdCloudflare WorkersGemini (AI Studio)OptunaNVIDIA cuDF / RAPIDSBigQueryThree.js / WebGL
Live siteCodePress

Core repo private — the linked repo is the public explanatory layer.

Lore

Who Knows What in an Open-Source Project

2026 · Graph Hacks (FalkorDB × WeMakeDevs) · submitted Oct 18

Who knows this part of the code, and who should look at this issue? Lore reads an open-source project’s public GitHub history into a FalkorDB graph: who wrote and reviewed which PRs, which files each one changed, which issues it fixed. A Gemini agent answers from that graph with twelve Cypher tools, and every answer cites the PRs behind it.

While it works, an Under the hood panel shows each tool it picks and each Cypher query it runs, then draws the evidence as a graph. Triage finds related issues and an active reviewer for a new issue and drafts a reply without posting it. Teams runs label propagation over who reviews whom. Signed in with GitHub, each person gets a memory graph of their own, and anyone can add a public repo.

Backtested on 300 real FalkorDB PRs: Lore had a real reviewer in its top three 90% of the time, against 88% for “the most frequent reviewer”. On the single top pick that baseline wins overall, and the write-up says so.

The live demo runs on free tiers, so the first load can take up to a minute.

1,357FalkorDB PRs in the graph
90%real reviewer in the top 3
12Cypher tools for the agent
₹0to host, all free tiers
FalkorDBCypherGraph AlgorithmsAI AgentsGemini APIMCPFastAPIGitHub ActionsDocker on RenderFalkorDB Cloud
Live demoCodeDemo videoBlog post

Wiretrap

Adversarial Security Testing for AI Agents

2026 · Google Cloud Gen AI Academy APAC — Cohort 3

“We don’t grade what your agent says. We audit what it does.” Most LLM security tools read the agent’s reply and grade the text. This one reads the tool-call log.

Every tool the agent declares gets cloned as a dead copy with the same name and JSON Schema. The copy records the call and returns fake data, so nothing real fires. Attacks run in separate Gemini conversations and a judge model reads the log. The payloads are written to look like ordinary work, no shouting and no “ignore all previous instructions”, because current models refuse those easily.

It caught a Google codelab agent saying “I don’t have access or authorization to modify configuration files” after it had already run ls -la through its shell tool. Text scanners pass that. Wiretrap calls it “Said No, Did Yes” and shows the arguments.

A run with nothing classified reports INCONCLUSIVE, never HARDENED. Attacks against tools the agent doesn’t have are dropped from the score. Approved fixes get re-run against the attacks that broke it, and the report says when one didn’t hold.

The live demo runs on free tiers, so the first load can take up to a minute while it wakes up.

8attack categories
9.4Klines of TypeScript
₹0to host, all free tiers
4model fallback ladder
Adversarial TestingAgent SecurityGemini APIReact + TypeScriptExpressFirestoreFirebase AuthDocker on RenderGoogle AI Studio
Live demoCode

VerseCanvas

AI Poetry-to-Art Generator

2025 · Personal project

Turns a poem into original artwork through a multi-stage pipeline: semantic analysis with Gemini 2.0 → automated prompt engineering → image synthesis with Imagen 3.0 on Vertex AI → hybrid editing and text overlay. Six languages, six art styles, tunable mood intensity.

Gemini 2.0Imagen 3.0Vertex AIStreamlit
Live demoCode
Seeking grants & partners

SmartDesk

Multi-Agent Assistant with Measured RAG

2026 · Gen AI Academy APAC — Cohort 1 Hackathon · top 101 of 1,500+

A root ADK orchestrator routing to specialised sub-agents — InboxAgent, PlannerAgent and a DataAgent doing retrieval over a pgvector knowledge base — with a SequentialAgent synthesising the reply. Gmail and Calendar are reached through MCP servers I wrote myself against the protocol SDK over stdio, not off-the-shelf toolsets. Containerised on Cloud Run.

The retrieval is evaluated, not assumed: 40 labelled questions over 120 notes, recall@k and MRR@k, significance by paired bootstrap at 10,000 resamples. That harness is what stopped a bad decision — a cross-encoder reranker looked like a clear win on the development embedder (MRR@10 +0.094, p=0.04), then collapsed on the production embedder to +0.028 with a CI spanning zero. Shipping it would have added 190 MB of PyTorch and taken search from 29 ms to 2,382 ms, in an image deliberately cut from 1.8 GB to 340 MB, for a gain indistinguishable from noise. The baseline ships instead, now on evidence rather than caution.

0.963recall@5
40labelled questions
82×latency the reranker cost
54tests
Google ADKMCP (servers written from scratch)pgvectorRAG evaluationPaired bootstrapCross-encoder / RRF / BM25Gemini 2.5 FlashCloud Run
CodeEval results

Location Intelligence Agent

MCP over BigQuery + Google Maps

2026 · Google Cloud Gen AI Academy APAC

An ADK agent wired to two remote MCP servers — BigQuery for demographic, pricing and sales data, Google Maps for location analysis. Scores zip codes by demographics and foot traffic, flags underperforming stores, projects next-month revenue and measures competitor density.

MCP ServersBigQueryGoogle Maps APIADK
Code

Smart Money Tracker

🏆 OpenServ × Hack2skill

Multi-Agent Blockchain Analytics

2024 · Winner — Best DeFi Application

Multi-agent whale tracking across Ethereum, BNB Smart Chain and Polygon using the OpenServ SDK, with Alchemy webhooks driving sub-second transaction monitoring. Confidence-scored detection of accumulation, distribution and coordinated movements, with automated exchange-wallet exclusion.

OpenServ SDKAlchemyMoralisCloud Runpgvector

SMT Chain Adapters

Mantle & BNB hackathon builds

2026 · Hackathon adapters

Thin per-chain adapters that import the SMT brain and expose it to a specific ecosystem — keeping one shared decision engine behind platform-specific front ends rather than forking the strategy code.

MantleBNB ChainPython
LiveMantleBNB
Seeking grants & partners

AutoKorrekt

AI Answer-Evaluation Platform for Teachers

May 2024 – Jan 2025 · Lead Frontend Developer · EdTech startup MVP

An evaluation platform that takes a teacher from sign-up to graded scripts: create a class, upload a student roster and question paper, bulk-upload scanned answer sheets, then track each student through OCR, AI evaluation and validation. I led the frontend and designed the ML data flow across S3, RDS and SageMaker — interactive PDF processing with coordinate-based text extraction over AWS Textract, a bilingual EN/DE feedback interface on Amazon Comprehend, and a per-student status pipeline that scaled to 1,000+ concurrent submissions.

Teacher sign-upTeacher sign-upSign inSign inCreate a class — student roster (CSV) + question-paper uploadCreate a class — student roster (CSV) + question-paper uploadPer-student evaluation pipeline & bulk answer-sheet uploadPer-student evaluation pipeline & bulk answer-sheet upload
ReactNext.jsTypeScriptAWS TextractAmazon ComprehendSageMakerDjango

Asha Chatbot

JobsForHer Foundation

2025 · Hackathon project

A context-aware chatbot surfacing job listings, community events and mentorship programmes, and handling FAQs — built to support women returning to and advancing in the workforce.

Conversational AINLPPython
Code

MedTestAI

HIPAA-Compliant Healthcare AI

2025 – 2026 · Personal project

A healthcare AI project built around HIPAA compliance constraints — handling protected health information under the access, auditing and data-handling rules regulated medical software requires.

Healthcare AIHIPAAJavaScript
Code

WebContentQnA

Ask questions about any page

2025 · Personal project

Paste a set of URLs and ask questions about their contents; the app retrieves and answers from page text using a DistilBERT question-answering model.

DistilBERTQuestion AnsweringTransformers
Code

Automotive Surveillance System

Capstone — Computer Vision

Dec 2024 – Jan 2025 · Great Learning

Vehicle make/model/year classification reaching 69.2% accuracy with ResNet50, plus Fast R-CNN object detection across 16,185 images spanning 196 classes.

ResNet50Fast R-CNNTensorFlowOpenCV
Code

MultiModel Botanical Classification

Plant species recognition

Jun – Jul 2024 · UT Austin McCombs

83.79% accuracy with a CNN for automated plant-species classification across 4,750+ seedling images, benchmarked against Random Forest and SVM baselines.

CNNTransfer LearningComputer Vision
Code

Signal & SVHN Classification

Neural networks, two domains

Aug – Sep 2024 · UT Austin McCombs

Neural networks reaching 90% accuracy on signal-quality prediction and 85% on SVHN street-number digit recognition, with regularisation and a visualisation suite for real-world noise.

Deep LearningKerasRegularisation
Code

Customer Churn Prediction

Telecom retention

Oct – Nov 2024 · UT Austin McCombs

81.48% accuracy identifying customers likely to cancel, comparing Bagging, AdaBoost and Gradient Boosting, and converting the model into a prioritised retention strategy.

Ensemble Methodsscikit-learnBusiness Strategy
Code

Semiconductor Yield Prediction

591 sensor signals

Sep – Oct 2024 · UT Austin McCombs

Pass/fail yield prediction from manufacturing sensor data, using feature selection and SMOTE to cut 591 signals down to the ones that actually carry information about defects.

Feature SelectionSMOTEQuality Control
Code

YouTube AdView Analytics

Ad-revenue prediction

Jul – Aug 2024 · Internship Studio

Random Forest regression predicting advertisement views from engagement metrics across 15,000+ videos, packaged for integration with business dashboards.

Random ForestpandasPredictive Analytics
Code

NBFC Loan Default Prediction

Credit risk

Feb 2025 · Great Learning Hackathon

A classification model predicting whether a client will default on loan repayment, built against a non-banking financial company dataset under hackathon time constraints.

ClassificationRisk Modelling
Code

Automobile Analysis & Classification

K-Means + PCA-enhanced SVM

Apr – May 2024 · UT Austin McCombs

Vehicle segmentation with K-Means clustering and classification with a PCA-enhanced SVM, analysing MPG, displacement, horsepower and weight across 102 features.

K-MeansPCASVM
Code

Stock Market Sentiment Analysis

Financial news NLP

Nov – Dec 2024 · UT Austin McCombs

An ensemble sentiment classifier over financial news using Word2Vec, GloVe and SBERT embeddings, with weekly market summaries generated by a Mistral-7B LLM.

NLPSBERTMistral-7BTransformers
Code

New-Wheels Performance Analytics

SQL business intelligence

Jul – Aug 2024 · UT Austin McCombs

Customer behaviour and sales analysis for a vehicle-resale business covering 994 customers and $1.25B revenue, with a KPI tracking system identifying key markets.

SQLBusiness IntelligenceKPI Design
Code

Business Analytics Portfolio

Pricing + HR attrition

May – Jun 2024 · UT Austin McCombs

Two supervised-learning studies: 86% accuracy on used-car price prediction for Cars4U, and 84% on employee attrition for a healthcare HR dataset.

RegressionHR Analyticsscikit-learn
Code

E-commerce Behaviour Dashboard

User & sales insight

2024 – 2025 · Personal project

A dashboard surfacing user-behaviour and sales-trend insights for an e-commerce site, turning raw event and transaction data into merchandising decisions.

DashboardsAnalyticsPython
Code

Applied Statistics & Hypothesis Testing

Statistical foundations

May – Jun 2024 · UT Austin McCombs

Hypothesis testing on car-purchase behaviour, manufacturing quality control identifying a 5% defect rate, and probability models for weekly sales forecasting.

Hypothesis TestingSciPyProbability
Code

Yatra Test Automation Framework

Hybrid Selenium framework

2024 · Personal project

A hybrid data-driven Selenium framework with Page Object Model structure, external test data, cross-browser runs on Edge/Chrome/Firefox, HTML reporting and failure screenshots — wired into Jenkins.

SeleniumPyTestPOMJenkins
Code
Claude CodeAI-Native CI/CDPythonGoogle ADKMCPVertex AIGeminiImagen BigQueryAlloyDBpgvectorCloud RunCloudflare WorkersOracle CloudGoogle AI Studio NVIDIA cuDFRAPIDSCUDAThree.jsPub/Sub TensorFlowPyTorchscikit-learnOptunaCatBoost FastAPIStreamlitReactTypeScriptDockersystemdSQL Claude CodeAI-Native CI/CDPythonGoogle ADKMCPVertex AIGeminiImagen BigQueryAlloyDBpgvectorCloud RunCloudflare WorkersOracle CloudGoogle AI Studio NVIDIA cuDFRAPIDSCUDAThree.jsPub/Sub TensorFlowPyTorchscikit-learnOptunaCatBoost FastAPIStreamlitReactTypeScriptDockersystemdSQL
About

From breaking systems
to architecting them.

I spent four years at Deloitte leading a QA automation team — finding the ways enterprise systems fail before customers did. That work taught me something most ML curricula skip: a model that cannot tell you when it is wrong is not finished.

So when I moved into AI, I built the way I used to test. My flagship system refuses to trade on a data feed it cannot verify, logs a plain-English reason for every decision, and gates every weekly model update behind statistical checks built to catch overfitting. When I graded its six AI agents on independent data, all six called direction at a coin flip, so I rebuilt it: written rules with replayed records open trades, and the agents became advisors graded on every call.

The way I work has changed too. I don't hand-type a 91,000-line platform. I specify it, then direct AI coding agents to implement it. Claude Code in a terminal is my build environment: it generates targeted hotfixes and structures the pull requests, and nothing merges until I have audited the diff myself. The speed-up is real and so is the gate. I am the only one standing between a generated patch and a system trading live.

That makes operations the other half of the job. The platform runs 24/7 as a daemon on an Oracle Cloud free-tier VM (I moved it off GCP in two days when the billing account closed), and I read its logs the way I used to read defect reports, triaging execution errors and catching systemic drift before it costs anything. The no-silent-fallback rule I designed in has already surfaced four multi-week data-integrity bugs.

I completed my PGP in AI/ML at UT Austin McCombs & Great Lakes at Rank 1, then Google Cloud's Gen AI Academy, covering Vertex AI, Gemini, ADK, MCP and AlloyDB. All of it is architected and operated from a single 16 GB laptop; when it needed to go faster, a cuDF/RAPIDS rewrite cut the validation pipeline ~6× on an NVIDIA T4. I built the whole portfolio while a full-time family carer. Today I'm looking for teams building agentic systems that have to survive contact with the real world.

⚙️
Provisional patent filed

India, App. No. 202631090789 — the SMT multi-agent decision architecture.

📈
WEEX AI Wars I — #1 preliminary leaderboard

Top of 230+ teams, then one of 37 finalists in the $880K live AI trading competition. Coverage

🏆
Winner — Best DeFi Application

OpenServ × Hack2skill. Runner-up for Best Website Application.

☁️
Top 101 of 1,500+ teams

Google Cloud Gen AI Academy APAC 2026, Cohort 1 hackathon, with SmartDesk — advanced to prototype refinement.

🥇
Rank 1 in batch

PGP in AI/ML, UT Austin McCombs & Great Lakes — GPA 4.09/5.

📜
12 credentials, all verifiable

Google Cloud Gen AI Academy (2025 & APAC 2026 Cohorts 1–3) · Gen AI Exchange · Agentic AI Day · Asha AI Hackathon · CPSAT. See them all

Experience

Where I've worked.

Founder & sole engineer

JTech Solutions (pre-incorporation) · Smart Money Trading (SMT) · patent pending
Jan 2024 – Present

Building SMT, an AI trading agent that has to prove why it traded, and SMT World, the product exchanges give their users on top of it. 91,000 lines of Python, 3,045 tests, 344 merged pull requests and 611 measured findings, written by AI coding agents I direct, with me as the only review before anything merges. Graded my own six AI agents on independent data, found all six at a coin flip, and rebuilt the system so written rules open trades and the agents are graded advisors. Tested 1,800 strategy variants (36 million simulated trades) at six exchanges’ fees. Audited 98 AI-written merges and found an order bug that had rejected 1,137 live orders over three months while every test passed. It places real orders 24/7 on WEEX’s simulated account from an Oracle Cloud free-tier VM, moved off GCP in two days. Began at the OpenServ × Hack2skill hackathon, where the predecessor build won Best DeFi Application; later #1 on the WEEX AI Wars I preliminary leaderboard of 230+ teams. All while a full-time family carer.

Lead Frontend Developer

AutoKorrekt — EdTech startup MVP
May 2024 – Jan 2025

Interactive PDF processing with coordinate-based text extraction, integrating AWS Textract OCR and Mistral models into an AI answer-evaluation workflow. Bilingual EN/DE interface scaling to 1,000+ concurrent submissions.

Machine Learning Intern

Internship Studio
Jul 2024 – Aug 2024

Random Forest model predicting YouTube ad views across a 15,000+ video dataset, packaged for integration with business dashboards.

Test Automation Engineer (Analyst)

Deloitte Consulting LLP — Bangalore
Aug 2019 – Jan 2023

Led a 6-member QA team for Fortune 500 clients including AT&T and Hewlett Packard Enterprise. Improved test execution efficiency 83%, tracked quality across 343 components, and analysed 50,000+ SAP Hybris transactions in Python.

QA Automation Intern

Deloitte Consulting LLP · Bangalore
Jan 2019 – Jul 2019

Functional testing of the North Dakota State Government Health and Human Services website, plus POEDI XML script testing. Ran environment testing loops, tracked the active regression suites, and managed system errors in Jira. Converted to the analyst role above after six months.

Credentials

Every certificate, verifiable.

Badge profiles first — they are confirmable by a third party and stay current. The certificate files are linked underneath for anything the profiles don't cover.

Credly badgesGoogle Skills profileAll certificate files
🎓

Post Graduate Program in AI & Machine Learning

Texas McCombs School of Business & Great Lakes · Mar 2025

GPA 4.09/5 · Rank 1 in batch

CertificateGrade bookVerify
☁️

Gen AI Academy — Completion

Google Cloud & Hack2skill · 28 May 2025

Vertex AI · Gemini APIs · Imagen · Streamlit · Multimodal RAG

CertificateID 2025H2S04GENAI-A01917
☁️

Gen AI Academy APAC 2026 — Cohort 3, Completion

Google Cloud & Hack2skill · Sep 2026

Agentic AI on Google Cloud — the cohort Wiretrap was built in

☁️

Gen AI Academy APAC 2026 — Cohort 2, Completion

Google Cloud & Hack2skill · 31 Jul 2026

Data, analytics and intelligent systems on Google Cloud

CertificateID 2026H2S07GCGENAIAPACC2-P02585
☁️

Gen AI Academy APAC 2026 — Cohort 1, Completion

Google Cloud & Hack2skill · 28 Apr 2026

Building, connecting and migrating AI-powered applications

CertificateID 2026H2S04GCGENAIAPACC1-P01679
☁️

Gen AI Academy APAC 2026 — Cohort 1 Hackathon

Google Cloud & Hack2skill · 19 May 2026

Shortlisted top 101 of 1,500+ teams

CertificateID 2026H2S05GCGENAIAPACC1-P00373
☁️

Gen AI Exchange Hackathon 2025

Google Cloud & Hack2skill · 14 Jan 2026

Problem statement — “Automating Test Case Generation with AI”

CertificateID 2025H2S08GH-P100275
☁️

Agentic AI Day

Google Cloud & Hack2skill · 2025

Build the next generation of intelligent agents

CertificateID 2025H2S06AID-I01369
🏆

Asha AI Hackathon 2025

JobsForHer Foundation & Hack2skill · 2025

The hackathon behind the Asha Chatbot project

CertificateID 2025H2S03AAIH-P00105
📜

Machine Learning Internship

Internship Studio · Jul – Aug 2024

YouTube ad-view analytics

InternshipTrainingID ISMAIL600552
📜

AI-Preneur Workshop

AI-Preneur · Mar 2025

Freelancing with AI — business applications

Certificate
📜

CPSAT — Certified Professional Selenium Automation Testing

CPSAT · 2022

Backing & collaboration

Actively seeking grants and partners.

SMT is patent-pending, running live 24/7, and looking for its first backer. If you fund early-stage AI, run a grant programme, or want to build on it — leave your details and I'll come back to you.

Fund, partner, or just say hello.

I'm looking for grants and early-stage funding for Smart Money Trading, and I'm open to collaboration on any of the projects here. Tell me who you are and I'll follow up personally.

For grants and funding specifically, Telegram, Discord or jtech26smt@gmail.com reach me fastest. The company on one page: founder résumé (PDF).

jtech26smt@gmail.comTelegramWhatsApp

Let's build something
that actually ships.

Open to applied AI engineering, agent evaluation and agent security roles, remote. The inbox is always open.

Talk to me

jannetekka96@gmail.comWhatsApp+919078802572Telegram

Find me

LinkedInGitHub@EkkaJanny96@JTechSMT (SMT)

Hire me

HimalayasLinkedIn