Which chunking works?
DocuMind · 35 answerable questions, same context size
Share of questions where the right passage ranks first. One question is worth 2.9 points, so gaps of one or two questions are noise.
( About )
0students led through a STEM-career awareness project
0internships and traineeships: IoT and robotics, machine learning
0projects built: an AI workspace, a payments engine, a RAG pipeline and an agent gate, each with tests
( Selected work )
Work
( Evidence )
Measured, not claimed.
Each project started from a test set. These are the real numbers, with the catch next to each one. I wrote every test set myself, so they show a design choice paid off, not how it would score in the wild.
Does the rule engine hold up?
RecoverFlow · 47 hard payment-failure messages
Diagnosis accuracy, with 0 unsafe actions. I fixed the rules using these same cases, so this is a regression gate, not live accuracy.
Does the gate stop attacks?
AgentGate · 44 attacks and 12 legitimate calls
Denied or allowedHeld for a human
37 attacks denied, 7 held for approval, none executed. 11 legitimate calls allowed, 1 held by design, none wrongly denied. Scripted agents and simulated tools.
Where it is still weak
DocuMind · answering without a language model
It never answered a question the docs don't cover, but it answers fewer than a third of the covered ones correctly. Gemini-based answers are the next measurement.
( Try it )
Try to break the gate.
This is AgentGate's real policy engine and argument checks, copied from the repo and running in your browser against a simulated workspace. Nothing is executed. The challenge: get ₹15,000 of refunds through without a human.
1 · Pick a tool call
2 · What the gate decides
Run so far
Pick an example or write your own call, then send it. Try the ones that look sneaky.
3 · The policy it is judging against (editable JSON)
- Nothing sent yet.
( Toolkit )
Toolkit
Languages
C++PythonTypeScriptJavaScript
Frontend
ReactNext.jsTailwind CSSTipTap / ProseMirrorFramer Motion
Backend
Node.jsExpressMongoDB AtlasMongooseJWTSSERazorpay webhooks
AI
Gemini APILLM streamingPrompt engineeringMachine learning
Tools
GitGitHub ActionsDockerVercelRender
Hardware
IoTSensorsHardware integration
( Path )
NOW IMPROVING
RecoverFlow, DocuMind and AgentGate
RecoverFlow, DocuMind and AgentGate are live. AgentGate's site and playground run the real policy engine; its approval dashboard still needs a server I have not hosted.
Next: host the dashboard, record demo videos, and run the model-based evaluations.
Path
2025 – 2029
B.Tech, CS (AI/ML)
NxtWave Institute of Advanced Technologies (NIAT). Foundations in C++, Python, data structures and algorithms, and machine learning.
Aug – Sep 2024
T-Works · IoT & Robotics Trainee
Technical engineer on AI-powered smart street lights: hardware integration, sensor calibration, IoT connectivity, and the link between the AI model and the physical system.
Aug 2024
Rubaroo · Community Volunteer
Created and led a small STEM-career awareness project that engaged 100+ students through discussion, storytelling and interactive sessions.
Apr – Jul 2024
Airbac Labs · Student Intern
Learned Python and machine learning and applied them to practical projects, including stock-market prediction models.
( How I build )
How I build
I build with AI coding tools like a fast junior engineer: I write the spec, review every change and own the decisions. Then I read the code, run the tests and trace the key flows until I can explain how each part works.
( Contact )
Contact
Let's talk.upparikarthik505@gmail.com · Hyderabad, India