currently training:
□ MTech Data Science · epoch 1/4
□ Cachy · knowledge engine
□ RAG evals · reading list
▊
AI engineer in training. I build models, ship the apps around them, and keep notes on what didn't work.
13 projects shipped · 2 papers under review · 4 experiments running
// training progress
Self-scored. Each bar lists the shipped projects it comes from.
Primary language of every shipped AI project.
✓ Cachy ✓ AirSwipe ✓ IPL pipeline ✓ +10 more
Two research papers built on it.
✓ ProbCLIP-A ✓ preference paper
MediaPipe, OpenCV, CLIP in shipped products.
✓ AirSwipe ✓ Cachy OCR ✓ ProbCLIP-A
Built retrieval from scratch; still the newest skill here.
✓ Constitution QA ✓ Cachy LLM chain → current focus: evals
pandas, SQLite, messy real-world Excel included.
✓ Bangalore AQI ✓ 14-station consolidation
Drift detection, monitoring, multi-service deploys.
✓ IPL drift pipeline ✓ HF Spaces deploys
Most of my models end up inside an app.
✓ 7 shipped apps ✓ Flutter ✓ Swift
// experiment board
Failed experiments stay on the board.
// 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.
// pinned to the desk
Each card shows the pipeline it ran through. All code on GitHub.
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.
Whisper · OpenCV · Tesseract · SQLite · HF Spaces
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.
sentence-transformers · ChromaDB · Mistral-7B
flip →
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.
XGBoost · FastAPI · Streamlit · Docker
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.
MediaPipe · OpenCV · PyQt6
data mining · unsupervised
A year of data from 14 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.
pandas · scikit-learn · DBSCAN
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.
ML Kit · Flutter · Kotlin IME
// the app shelf
Five native macOS apps in Swift, plus mobile and web. Each one started as something I wanted on my own machine.
Keeps a Mac awake, lid closed included. Timed durations, global shortcut, and smart triggers: auto-enables on chosen apps, Wi-Fi networks, high CPU, or active downloads. Scriptable via insomniac:// URL scheme.
Swift 5.9 · AppKit · IOKit · NWPathMonitor
Custom 3/4/5-finger trackpad gestures for window snapping, media control, launching apps. A slow swipe switches windows, a fast flick opens the browser. Reciprocal undo, haptic feedback, per-app filters.
Swift · AppKit · IOKit multitouch
Themes the macOS menu bar, which Apple gives you no API for. Click-through overlay at the menu bar's own window level, gradients and blur, auto-presets on window state, notch-aware, and it reads menu titles via Accessibility API to keep them legible.
Swift · SwiftUI · Accessibility API · CGWindowList
Dims displays below the hardware minimum using overlay windows, each connected monitor independently. Lives in the menu bar, does one thing well.
Swift · SwiftUI · AppKit
Your own photos as desktop widgets: four sizes, per-widget photo choice, full color even in monochrome widget mode.
Swift · WidgetKit · AppIntents
A closet manager: photograph your clothes once and the app plans what you wear.
Flutter · SQLite · on-device, no account
A career dashboard answering one question each morning: what should I work on today to get hired?
Next.js 16 · React 19 · TypeScript · Pyodide
// running total: 13 projects, all listed on this page.
// research shelf
Hover a book to see the projects behind it.
// research wall · click a paper
Written as an MTech student, trained on free GPUs. Click to flip.
status: under review
Siamese DeBERTa with swap augmentation & temperature calibration
trained for $0 !flip →
Predicts which chatbot answer humans prefer, at 98% of state-of-the-art with a model 127× smaller (71M params), trained for $0 on free T4s in 8.4 hours.
log loss 1.052 → 0.987 · accuracy +3.5pts
what I learned: swap augmentation bought more accuracy than a bigger model would have.
status: under review
Probabilistic adapters for frozen CLIP vision foundation models
knows when it's unsureflip →
Makes CLIP say how sure it is: a 4.2M-param adapter on frozen CLIP outputs distributions, Monte Carlo sampling turns them into uncertainty scores that catch 36.6% of failures at 8.1% false alarms.
R@1 68.9% · ECE 0.062 (best of all baselines)
what I learned: skip KL warmup and the variance collapses to zero.
// 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 first-author papers trained entirely on free-tier GPUs. Both repos document 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.
// git log --journey
// open a pull request
Looking for AI/ML internships and roles.
end of notebook, for now.
