The journey so far
I'm Aditi Singla, finishing an Integrated B.Tech + M.Tech in Computer Science at IIIT-Delhi. This page is the honest version of the last five years: what I built, what broke, and what each thing taught me that the next thing depended on.
The setup
Joined the Integrated B.Tech + M.Tech program in Computer Science at IIIT-Delhi. Five years, two degrees, one transcript. The deal with myself was simple: leave with things that actually run, not just grades. (The grades held up too: 9.29/10.)
Foundations, the unglamorous way
First repositories were C programs for operating systems and computer organization: process scheduling, memory, assembly. Nothing looked like much at the time. It turns out that year decided how I debug everything since: read the source, trace the syscall, check the actual state.
Data that has to be trusted
Grad-Guru merged ten scraped datasets with conflicting schemas into one 100,000+ record PostgreSQL warehouse, and cut manual college-research effort by about 90% for the people who used it. Same year, the opposite end of the stack: a two-pass assembler for a custom 16-bit ISA, where a missing HLT instruction is a hard error, not a style opinion.
Down to the metal
Built a UNIX shell from scratch: fork and exec dispatch, readline history, reaping children, threading. Four of us, one semester, and the discovery that everything a shell does looks trivial until it is your fork() call hanging. Around then the institutional work started finding me, and I learned that tools people rely on weekly need tests, not just features.
The year everything shipped
NeuroSync: a PyTorch network fusing facial and vocal emotion signals for real-time stress detection, hitting MSE 0.0032 with chat and VR front-ends. The Graduation Checklist Portal that replaced an academic team's Excel macros. A Django portal unifying 17+ record APIs. A RAG chatbot that answers only from your documents. Different problems, one lesson: the fallback path is the product.
Agents that prove their work
The IneqMath Proof Agent: an LLM that generates inequality proofs and gets them type-checked in Lean 4, with GEPA prompt optimization raising the verified pass rate from 28.6% to 100%. Then Paper2Beamer, an agent that turns papers into presentation decks. Graduation is next; the job search is on, and the receipts are public.
Every number on this site traces to a public repository: the metrics, the traces, the results directories.
APIs fail, quotas run out, repos get deleted. NURO answers without its LLM; the RAG service runs on a free model; the crawler respects rate limits by construction.
LICENSE files, CI pipelines, test suites, .gitignore hygiene. That unglamorous layer is the difference between a demo and production-grade work.