▊
AI engineer in training. I build models, ship the apps around them, and keep notes on what didn't work.
19 projects shipped · 2 papers under review · 2 building next
// these six run right now, in this tab
press Ctrl ` for a terminal, or
// pinned to the desk
Each card shows the pipeline it ran through. All code on GitHub.
the running web app
knowledge engine · flutter + fastapi
Turns Reels, Shorts and articles into structured knowledge cards: transcription, OCR, an LLM chain with automatic fallback across Gemini / Cerebras / Groq, all linked in a semantic knowledge graph.
3-provider LLM fallbacksemantic graph1 SQLite, 0 Redis
Whisper · OpenCV · Tesseract · SQLite · HF Spaces
try the web appproduct siteandroid apk
rag from scratch · no langchain
Ask the Indian Constitution anything, get grounded answers with exact citations and similarity scores. Retrieval written from scratch in ~60 lines. A 20-query stress test is documented in the repo, 78% accuracy after fixing chunking.
78% on 20-query eval~60 linesno framework
sentence-transformers · ChromaDB · Mistral-7B
flip →
red circles are the hotspots
data mining · unsupervised
A year of data from 13 CPCB stations, 6 engineered features, three clustering algorithms compared. DBSCAN worked best since it flags outliers, and the outliers were the hotspots. Silk Board hit AQI 500.
13 stations · 1 year3 algorithms comparedhotspots = outliers
pandas · scikit-learn · DBSCAN
explore the findings
5 minhow often the monitor recomputes population stability index on a live XGBoost service
full ml lifecycle · self-monitoring
XGBoost match-outcome model served over FastAPI, with a monitor computing Population Stability Index every 5 minutes to catch data drift, and a live Streamlit health dashboard. Three services, Docker Compose.
PSI every 5 min3 serviceslive drift dashboard
XGBoost · FastAPI · Streamlit · Docker
6gestures, orientation-invariant, with dwell-time guards against accidental triggers
computer vision · gesture control
Control PowerPoint with bare hands through a webcam: swipe to navigate, point for a live laser dot, pinch to zoom. Orientation-invariant detection with dwell-time guards against accidental triggers.
orientation-invariantdwell-time guards6 gestures
MediaPipe · OpenCV · PyQt6
macOS app
15classes across 5 crops, every prediction paired with a Grad-CAM heatmap
cnn · grad-cam explainability
Upload a leaf photo, get a diagnosis, and a Grad-CAM heatmap showing exactly which part of the leaf the model looked at. Because a prediction plus a confidence score tells you nothing about whether it read the lesion or the background.
15 classes · 5 cropsGrad-CAM heatmapseverity levels
PyTorch · Grad-CAM · Gradio
100%on-device. no network call, no account, Catmull-Rom smoothing for hand tremor
on-device ml · accessibility
Handwriting-to-text Android keyboard for seniors: on-device ink recognition, a personal dictionary that learns your writing, and Catmull-Rom smoothing to filter hand tremor.
100% on-devicetremor smoothingreal Kotlin IME
ML Kit · Flutter · Kotlin IME
android apk
CLIPtext and image queries against one index, on a domain it was never fine-tuned for
clip · text + image retrieval
Find any Indian dish by describing it in plain English, or by uploading a photo of something similar. Built to test whether CLIP holds up on something domain-specific instead of the usual stock-photo demos. It does.
text + image queriesdomain CLIP testinterpretable ranks
CLIP · Gradio · HF Spaces
// research · both PDFs are on this page
Written as an MTech student, trained on free GPUs. You can read both of them here — the numbers below are the ones in the PDFs.
under review
Through Position Bias Mitigation and Architectural Symmetry
| validation log loss | 0.9871 | 6.2% better than the baselines |
|---|---|---|
| parameters | 71M | 127× fewer — the winning solutions ran 9B+ on 8×A100 |
| three-class accuracy | 52.06% | up 3.5 percentage points |
| training | 8.4 h | two free Kaggle T4s, 57,477 human preferences |
what I learned: swap augmentation bought more accuracy than a bigger model would have. It killed the position bias directly.
Code and weights stay unpublished until the review decision.
submitted to Elsevier
Uncertainty-Aware Cross-Modal Retrieval via Probabilistic Adapters for Frozen CLIP Vision Foundation Models
| text→image R@1 | 68.9% | on Flickr30K, just past CLIP-Adapter at 68.8% |
|---|---|---|
| calibration error | 0.062 | down from 0.078 on plain CLIP |
| failures flagged | 69.3% | while touching only 4.9% of the correct results |
| trainable params | 4.2M | 2.1M per modality; CLIP itself stays frozen |
what I learned: skip the KL warmup schedule and the variance collapses to zero. The model quietly stops being probabilistic.
My part: conceptualization, methodology, software, validation, review and editing. Code and weights stay unpublished until the review decision.
// competition · 2026
Akshara Foundation · Datathon 2026
Analysed arithmetic learning outcomes across 1.38 million Karnataka government school students. Built indices, OLS residuals, and a 7-screen Streamlit dashboard to argue the gap is skill-shaped, not geography-shaped.
Python · pandas · scipy · plotly · Streamlit · 1.38M records
// the app shelf
Five native macOS apps in Swift, plus mobile. Most of these started as simple interfaces for my ML projects.
Custom 3/4/5-finger trackpad gestures for macOS.
Swift · AppKit · IOKit multitouch
.dmg download
Your own photos as desktop widgets.
Swift · WidgetKit · AppIntents
.dmg download
Menu-bar control for the Background Sounds feature Apple buried in Accessibility.
Swift · AppKit · HearingUtilities (private)
.dmg download
Two players name words aimed at the middle of the last pair, until they match.
Flutter · Firestore · Firebase Hosting
play it liveandroid apk
Photograph your clothes once; the app plans what you wear.
Flutter · SQLite · on-device, no account
try it in the browser
Keeps a Mac awake, lid closed included. Triggers on app, Wi-Fi network, CPU load or active downloads.
Swift · AppKit · IOKit · NWPathMonitor
.dmg download
Dims displays below the hardware minimum, each connected monitor independently.
Swift · SwiftUI · AppKit
download app
Tidies a photo library and can undo every batch. One logged choke point is what makes Undo real.
Swift · SwiftUI · ffmpeg · Core Location
// running total: 19 shipped, 6 you can open right now, 5 more building. all on this page.
Three utilities that fix one annoyance each. Most exist because the OS said no.
Turn any Chrome extension into a working, signed Safari extension — no $99/year Apple Developer ID.
Shell · Xcode toolchain · WebExtensions
Sign any macOS .app locally with a free Apple certificate. Kills Gatekeeper friction in one
command.
Shell · codesign · Xcode CLT
An Obsidian plugin that auto-hides both sidebars in fullscreen, then reveals them when you hover the window edge.
TypeScript · Obsidian API
// git branch --unmerged
Two projects in active development. Neither product is finished, but the modeling core of each one is done, and both cores now run in your browser with nothing to install. They are both about getting a model to admit when it does not know. Sahayak stops answering when the retrieval score is too low. Oracle widens its prediction set when it has not seen enough days yet.
runs client-side via Pyodide
building · flagship · core runs live
grounded rag + rules engine
Grounded, cited answers about India's 3,000+ welfare schemes, plus a deterministic eligibility engine. Ask "I'm a Karnataka farmer with 2 acres — what do I get?" and get a checklist, not a keyword search. Abstains when it isn't sure. Retrieval core shipped — from-scratch hybrid retriever + rules engine + eval harness.
hybrid retrieval · BM25 + TF-IDF · rules engine
run the retriever in your browser
live demo ↗
walk-forward, scored one day at a time
building · core runs live
on-device · conformal calibration
Honest, calibrated predictions about your own behaviour — the shown 80% is right ~80% of the time. The hard part is meaningful ML on a tiny, noisy, single-person dataset without lying about confidence. Modeling core shipped — from-scratch logistic regression + split-conformal, backtested.
conformal prediction · Platt scaling
run the backtest in your browser
live demo ↗
// training progress
What I know, and what I've built with them. The shipped work under each one is the part that matters.
// professional track
Amar Infotech – Ahmedabad, Gujarat · 01/2025 to 06/2025
iTechBrains – Ahmedabad, India · 06/2024 to 07/2024
// academic background
M.S. Ramaiah University of Applied Sciences – Bengaluru, Karnataka · Expected 2027
SAL College of Engineering – Ahmedabad, Gujarat · 2021 to 2025
Certifications — five from IBM on Coursera. Each links to the PDF.
// index cards · how I work
Small working systems beat ambitious prototypes. Cachy launched with a three-provider fallback chain because free APIs kept dropping.
If I can't evaluate it, I can't improve it. Even the skills section on this page has numbers.
I write core pieces from scratch before reaching for a framework. That's how the 60-line retriever happened.
Dead ends get documented. FAILURES.md ships with the repo because I keep needing it later.
// git log --journey
// the short version
✓ models that ship
Notebook → Dockerized FastAPI service → drift monitoring. Already done end to end for the IPL pipeline: three services, PSI checks every five minutes.
✓ RAG from scratch
Retrieval, chunking, and evaluation written by hand and stress-tested, with the failure analysis published in the repo.
✓ research rigor on a budget
Two papers under review, both trained entirely on free-tier GPUs. Both write up what didn't work along the way.
✓ the whole product
If the model needs an app around it, I build that too, in Swift, Flutter, or on the web.
// open a pull request
Looking for AI/ML internships and roles.
end of notebook, for now.
