Personal & college
A NEET prep app, an iOS study companion, a work log, a home lab and a house design.
An all-in-one NEET-UG 2027 study app for one student: past papers question by question, syllabus tracking, a daily plan and an AI tutor.
- Tutor server live, and the app is on the student's tablet
- 2,420 questions cut as images from 13 past papers, so diagrams stay exact, each tagged by chapter
- Chapter practice, timed mocks on an NTA-style test screen and a review queue of wrong answers
- Written explanations, each checked against the original question image
- AI tutor on a home server: adds the NCERT text, asks Claude and streams the answer, with the key on the server and a monthly spend cap
- Password-protected progress page and twice-daily backups through a Cloudflare tunnel
- Tutor prompt tuned with an eval hill-climb against a baseline
A personal iOS app that runs my IIT Madras study plan: tap to start a block, a lock-screen countdown, and planned versus actual hours every week.
- In daily use since 28 Sep 2026; the sync server has been live since 12 Oct
- AlarmKit timers and Live Activities that break through silent mode and Focus
- A gradual wake-up ramp, wake log and wind-down
- Writes the plan to its own calendar and keeps it in sync
- Planned versus actual hours per course, exported for a weekly review
- Sync server in Swift that shares the app's core code, live on my home server behind a Cloudflare tunnel
A private work log for macOS: clock in and out, breaks, what you worked on, away time and git commits, exported as one HTML report.
- Encrypted SQLCipher database with the key in the login Keychain; no server and no sync
- Append-only SHA-256 hash chain over every change, so entries are voided or amended, never deleted
- Machine-detected time kept visually separate from anything typed by hand
- Report with a day timeline, project and app breakdown, commits and an integrity statement
Local smart-home control on a self-hosted Home Assistant, including an AC whose remote protocol I reverse-engineered.
- Decoded the Lloyd AC's 15-byte IR protocol, then wrote a custom climate integration that builds each packet on the fly, validated against 43 real captures
- Atomberg ceiling fan as a template fan over its NEC IR codes
- Circadian lighting with a follower bulb for the evenings
- Alexa voice control, local and free, through emulated Hue
A Linux home server running my services in Docker, published on subdomains of my own domain through Cloudflare tunnels with nothing port-forwarded.
- Home Assistant and the Companion sync server in containers, with the NEET tutor server built for the same setup
- Firewall closed to everything outside the local network
- This portfolio on GitHub Pages at smithmankar.co.in, with DNS on Cloudflare
A two-storey home with a terrace on a 30 x 50 ft plot, drafted floor by floor as code-generated plans and then modelled in Blender.
- Plans generated as SVG by Python from coordinate tables in feet, then rendered and checked in headless Chrome
- Each floor locked before work on the next began
- Services designed in from the start: one extract shaft, a ceiling plenum for ducts and wiring, and an AC plan per room
A factory planner for the game Satisfactory that reads your save file, solves production with linear programming, and lays out the plan as a graph and on the map.
- Save-file parser to read the current world
- Linear-programming solver for production rates
- Auto-laid-out production graph and a Leaflet map view
An LLM extraction pipeline that read all 1,431 chapters of a web novel and built a queryable record of who knows what, and when.
- 45,549 facts, 7,304 relationships and 1,912 aliases resolved to 4,386 distinct entities
- Runs on a local Qwen3-14B through LM Studio, or a cloud model by config
- Alias resolution folds a character's many names into one entity
Classification models that predict which employees will leave, trained on IBM's 35-attribute HR dataset.
- Logistic Regression baseline: 84% precision
- Decision Tree with GridSearchCV tuning
- Feature analysis on 35 HR attributes
- Confusion matrix and classification report
A dataset of IMDb's Top 250 films, scraped from the chart and enriched with OMDb metadata.
- Chart scraped with BeautifulSoup, then each title fetched from the OMDb API
- IMDb's weighted rating rebuilt from its public title and ratings files