Enterprise softwareBuilt with AILess costLess timeReal-timeOffline-firstFull-stackShipped, not slidesEnterprise softwareBuilt with AILess costLess timeReal-timeOffline-firstFull-stackShipped, not slides
Editor's letter. Most portfolios are templates. This one is a zine: two inks, real paper grain, every drawing hand-coded. It's how I like to build software too: on purpose, and nothing extra.
Have a look around. Try the console on p.20, and grab the resume on your way out.
— Siddharth
p.04 — The pitchWhy AI × me
The pitch
The pitch
Enterprise software used to take a big team and a long year. I bring it with AI, at less cost, in less time.
0
REST endpoints in TaskFlow
0
Automated tests in CrowdMate
0
ML models compared, honestly
0
Platforms: web, desktop, Android, iOS
01
Frame
Understand the business problem, the users and the roles before a single line.
02
Architect
Pick boring, proven pieces: auth, RBAC, data model, real-time where it pays.
03
Build × AI
AI writes the scaffolding; I own the logic, the reviews and the edge cases.
04
Prove
Tests, accessibility audits and CI, so it's shippable, not a demo.
p.06 — FeaturesSelected work, 2025–26
Features
Things I built
Thirteen projects, from a role-secured team platform to an offline file-transfer app, a flight recorder for outages, a pothole mapper that rides along on your commute, a gateway that cuts AI bills and a clipper that turns long videos into ready-to-post shorts. Every drawing on this page is hand-coded in two inks. Tap see it for the real screen.
fig. 01 — the board, live
01Flagship · enterpriseFull-stack · 2026
TaskFlow
A project & team platform that behaves like enterprise software, because it is.
Kanban, calendar, dashboards, reports, file sharing and real-time chat, all behind role-based access for Admins, Project Managers and Members.
47 REST endpoints, JWT auth, bcrypt and centralised RBAC
Socket.IO chat rooms, typing indicators and live board updates
AI insights: project health scoring, priority hints, workload balancing, LLM summaries
Angular 19Angular MaterialExpress 5MongoDBSocket.IOChart.jsClaude API
Fast file transfer between Android, iPhone and Windows, with no internet, no router and no compression.
Pick files, folders, photos or videos, show a QR code, and they beam straight across: in any direction between Android, iPhone and Windows. The sender makes its own private Wi-Fi, so it works on a train or a mountain. Original photos keep their quality, dates and location, and every byte is verified before it counts as done.
Own transfer protocol: parallel TLS streams, small files packed together, big files in 4–8 MB pieces
BLAKE3 on every piece and every file; only a damaged piece is re-sent, and a dropped connection resumes where it stopped
Tested with 1,003 files + 2.4 GB, two forced disconnects: 0 errors · ~118 MB/s encrypted
Native code on each platform: Kotlin Wi-Fi Direct, Swift peer-to-peer Wi-Fi between iPhones, C++/WinRT hotspot on Windows
One QR code for crowded places: virtual queues and lost-person safety tags.
Scan a QR, get a live token, go get chai. If a child or elder gets separated, whoever finds them scans their tag and the family is alerted instantly, without either side seeing a phone number.
A privacy-first digital twin that learns a student only from what they approve.
Built from scratch at GATEWAYS 2026. The twin asks questions out loud, turns answers into candidate facts behind a consent gate, simulates "what if" choices with real numbers and speaks its advice. I owned the front end: the living line-art twin, the visual data stories and the voice UI.
Consent-first loop: every fact is approved before the twin learns it
Animated twin that mouths speech, plus heatmaps, fidelity and what-if views
Predicts milk quality from seven lab readings, and refuses to fake its accuracy.
Move the sliders and the grade updates live with confidence and plain-language notes. The interesting part: 976 of the 1,059 rows were exact duplicates, so a normal split scored a fake 100%.
Compared 10 models three ways, including grouped CV on truly unseen samples
Honest best: ~88% on unseen data, instead of a leaky 100%
Flask app with accounts, saved history, confusion matrix and feature importance
07Native AndroidMCA mini project · with Megha Raykar
Habit Tracker
"Digital Sanctuary": a calm habit app with streaks, progress and a mood journal.
Daily focus ring, one-tap completion, current and longest streaks, weekly overview and a journal with moods, wrapped in soft-blue Material 3 with dark mode.
MVVM with ViewModels, repositories, Coroutines and Flow
Room database, DataStore preferences, onboarding and auth flow
Reads your rent agreement or offer letter and marks the lines to question before you sign.
Upload a PDF, even a phone scan. TermGuard splits it into clauses, checks each one against a written red-flag checklist, explains it in plain English, Hindi or Kannada, and lists what a fair agreement should say but yours doesn't. Click a finding and the exact line lights up on the page. It gives information, not legal advice.
Quote check: a finding is only shown if its quote is really in your document, so the AI can't make clauses up
A flight recorder for outages: what broke, in what order, and the change that probably did it.
BlackBox pulls in logs, Alertmanager alerts, deploys and Slack chat, squeezes tens of thousands of log lines into the moments that matter, ranks the changes that might have caused the outage, and drafts a blameless postmortem where every sentence links to its evidence. A built-in chaos lab breaks a demo shop on purpose so every incident has a known answer.
34,033 log lines → 23 timeline items with Drain template mining and z-score spike detection
Found the real trigger first in 49 of 50 injected incidents, each hiding two decoy deploys
Postmortem sentences cite event ids that are checked to exist: 100% evidence-linked
A free QuillBot-style writing suite that rewrites sentence by sentence and tells you why.
Paste a stiff or AI-sounding draft and pick one of ten modes, from Natural and Academic to Shorten and Custom. Each changed sentence is marked; click it to see the reason and keep your original if you prefer it. Also a grammar checker, a summarizer, a "My voice" mode that copies your own style, and a Word rewriter with change requests and version history.
Edits the .docx itself, not a text copy: headings, lists, tables and the references section stay untouched
Change requests like "make the Methodology more formal" only touch that section, and every version is kept
Free where QuillBot's free plan stops: 1,500 words vs 125, 10 modes vs 2, compare modes side by side
Ride with a phone on the handlebar, get back a map of every pothole on the road.
A fine-tuned YOLO11 model finds potholes in dashcam video, ByteTrack follows each one across frames so it is counted once, and the GPS track from the same ride puts it on the map with a size estimate and a blurred photo. Repeat rides merge into "seen 3×", and a PDF groups the worst potholes by road for the city corporation.
Fine-tuned YOLO11n reaches 0.62 mAP50 on held-out test photos, 0.86 recall on close potholes
Tracking turns 1,238 raw detections into 22 map dots with zero double counts on a ride with known answers
Caught that 57% of the dataset had near-twin photos and split by twin group so the test score is honest
A gateway that cuts a company's AI bill without touching its app.
Tokenomy sits between an app and its language models and speaks the OpenAI API, so switching to it is a one-line change. Repeated questions come back from a cache for free, reworded ones from a smart cache that compares meaning, and easy questions go to a small cheap model while hard ones go to a strong one. A dashboard shows spend, savings and every request.
Router picks the cheap or strong model with 96% accuracy on held-out prompts
Caught a trap: "to Delhi from Mumbai" and "from Delhi to Mumbai" score 0.992 similar; a key-facts guard cut wrong cache hits from 2 in 8 to 0
Works with the standard openai SDK, including streaming, with 25 offline tests and CI that builds the Docker image
fig. 13 — one long video in, ready-to-post shorts out
13AI · videoFastAPI + Whisper + Next.js · 2026
HypeSift AI
Paste a YouTube link or upload a long video, press AI Generate, get finished shorts ready to post.
HypeSift transcribes the video, measures loudness, motion and scene changes, and picks the moments that work on their own. Each clip is cut on a complete thought, reframed to vertical around the speaker's face, captioned word by word and given titles, descriptions and hashtags for six platforms. Every clip shows a 0–100 score with the reasons behind it, and a browser editor and a plain-English assistant let you change anything before you download.
A 2:53 video becomes 4 captioned, reframed shorts with post copy in 90 s on a laptop GPU
A YouTube link is checked in 2 s and a 4:27 video downloads in 1080p in 20 s; the speech model transcribes 115 s of audio in 4.5 s
69 tests, including full upload and YouTube-link → AI Generate → MP4 runs; scores explained across 7 dimensions; it costs $0 to run: every model is local
One page, no fluff. Profile, education, skills, internship, projects, soft skills and interests, with real numbers. Readable by humans and by ATS bots.
