A
Agents that don't make things up
Tool-calling agents whose facts come from a database or an API, plus guardrails that stop injected instructions.
ProofTripwire checks each risky tool call in ~0.2 s.
Mohamed Ibrahim A. “Ibu” · Software & AI engineer · Coimbatore, India
I build AI systems.
I keep the receipts.
I'm a final-year IT student shipping AI agents and serverless apps on AWS. Most portfolios ask you to take their word for it. This one prints a receipt for every claim, including the numbers that make me look worse.
How to read this page
Each of the 14 underlined claims on this page prints a receipt: the evidence, how it was measured, and the catch. Start with my best-looking number, R² 0.94. Hiring? Paste your job description and the printer itemizes what I can prove for it, and what I can't. Or ask the printer anything about me: it answers only with receipts.
Ibu's receipts
for the page you're reading
- Claims on this page
- 14
- Backed by public proof
- 9
- Corrected by my code
- 4
- Invented facts
- 0
Audited by you: 0 of 14
01
About me
I'm drawn to the gap between a demo and something you can trust.
I'm Ibu, a final-year B.Tech Information Technology student in Coimbatore, graduating in June 2027. I build AI agents and serverless apps on AWS, often in public and often against a weekend deadline.
My Supermarket agent is never allowed to invent a price: every number comes out of Postgres through a tool. This page holds my claims to the same rule. Each one links to the code, notebook or article behind it, and where there's no public proof, the receipt says so.
Writing these receipts caught four places where my own write-ups disagreed with my code. Those receipts print the corrected facts.
Off the clock I'm into quantum computing. I organized an IBM Qiskit Fall Fest workshop on quantum algorithms.
How I work
- Ground the model. Facts come from data and tools, never from the model's memory.
- Publish the number that hurts. An R² of 0.94 means little without the 47.93% beside it.
- Own the loop. Fifty lines I can debug beat a framework I can't.
02
Tech superpowers
Four things I'm reliably good at, each with proof attached.
B
Serverless on AWS
Event-driven backends with least-privilege IAM, defined as code instead of console clicks.
ProofThe Git Bard wakes up every Friday without a server.
C
Money-grade full-stack
Apps whose totals reconcile to the paisa: integer money, row locks, idempotent handlers, and tests that pin it all down.
ProofCelebrAI ships 86 tests, zero runtime dependencies.
D
Honest ML
Models tested on unseen, time-ordered data, with the unflattering metric printed next to the flattering one.
ProofCryptoPulse scores R² 0.94. Its receipt shows what that hides.
Also fluent in Python · JavaScript · TypeScript · SQL & PL/SQL · Docker · GitHub Actions · Google Cloud Run · Firebase
Studied, not yet shipped Java · C · C++ · Kubernetes · Azure · BigQuery · Snowflake
03
Projects & achievements
Six featured builds, the rest in brief, then the record.
Featured builds
Supermarket Ops Agent
A neighbourhood store run entirely from Telegram: billing, stock, credit ledgers (khata) and GST invoices, with Claude choosing from 30 typed tools backed by Postgres.
DocuMind AI
Upload a PDF and get a summary, grounded Q&A, entities, tone and a two-document comparison, built in a weekend on Amazon Bedrock and Lambda.
Source: DocuMind AI Write-up: DocuMind AI on AWS Builder Center
Tripwire
Before an agent sends or shares anything, a small model asks one question, “did the user actually ask for this?”, without ever reading the injected text. It answers in ~0.2 s.
Ghost Probe
On a versioned S3 bucket, “deleted” only adds a delete marker. This read-only tool lets an agent see every hidden version, and it turned a COMPLETE deletion verdict into INCOMPLETE.
The Git Bard
Every Friday, a Lambda gathers my week's commits and emails them back as a fantasy chronicle written by Amazon Nova. Bug fixes become slain bugbears.
CryptoPulse
A crypto forecasting platform that shows how far to trust each forecast. Its LSTM scores R² 0.94 on unseen data, which is less impressive than it sounds.
More builds
-
Voter Saathi
A non-partisan election guide in 12 Indian languages: a rule-based flow for structure, an LLM for free-form questions (free OpenRouter models first, Gemini as the fallback).Node.js · OpenRouter · Gemini · Docker · Cloud Run
Source: Voter Saathi -
CelebrAI
An event planner whose budget engine works in integer paise, so every allocation adds up exactly. An LLM suggests; a schema gate decides.Vanilla JS · LLM · Tests
Source: CelebrAI -
Tiny Tale Toy
A choose-your-own-adventure where every branch is generated live by a model on Bedrock, behind API Gateway and Lambda.API Gateway · Lambda · Bedrock · Amplify
Source: Tiny Tale Toy -
EcoTrack AI
A carbon-footprint tracker with a multi-step calculator and Gemini coaching.React · Express on Cloud Run · Firebase · Gemini
Source: EcoTrack AI -
UniLeave & GrievanceHub
Two campus workflow apps for leave approvals and grievance tracking, with role-based modules and dashboards.Oracle APEX · Oracle Database · PL/SQL
Source: UniLeave -
Crisis Response Coordinator
A RAG prototype that retrieves the relevant safety protocol before it drafts guidance, for a human responder to act on. Code not public.Python · LangChain · FastAPI · Docker
Demo video: Crisis Response Coordinator -
Bone density classification
Knee X-rays sorted into normal, osteopenia and osteoporosis with fine-tuned Xception and MobileNet, plus Grad-CAM to check what the network looked at. No accuracy quoted until the evaluation is public.CNNs · Transfer learning · Grad-CAM
Code not public -
Voice screening UI
A multilingual interface for an Alzheimer's voice-analysis workflow. The analysis layer is still simulated; no model has been trained yet.React · TypeScript · i18n
Source: Voice screening UI
Achievements
-
Challenge entries
Three AWS Builder Center challenges: the Weekend Showcase (DocuMind AI, Aug 2026), “Set Your Creative App Free” (The Git Bard) and “Give your Agent a Tool” (Ghost Probe, Oct 2026). Plus HackerRank Orchestrate rounds, Aug–Oct 2026.
-
Experience
Three internships: AI Intern at Infosys Springboard (Batch 5), Backend AI Engineer Intern at FlyRank AI, and MERN Stack Developer Intern in the AICTE program.
-
Community
AWS Student Builder Campus Leader and Microsoft Learn Student Ambassador (Associate). I write on AWS Builder Center.
-
Education
B.Tech, Information Technology, Coimbatore. Graduating June 2027.
04
Contact
Hiring for software or AI engineering, or have a hard problem worth prototyping? Email is fastest.
See what I can prove for your role before you write.
A quick question first? It quotes this page, word for word.
This site's source github.com/MdIbuA/ibu-receipts