7+
products & research engines built
Proof of Work
Falsifiability-first research and products — where Sanskrit computation meets applied AI. A working catalogue of what I ship, and what I can prove.
Sparsh Sharma · Independent researcher · Melbourne
By the numbers
The Works
A standing record of what I ship and research. Some run today; some are still under the bench. Nothing here was not built by me — and nothing leaves without surviving falsification.
Run a commercial cleaning business from one screen — sites, contractors, clean records and risk alerts, all in sync. No more spreadsheets and lost message threads.
Open the app →A calibrated falsification harness for retrieval & ranking — a four-null gate plus an integrity lock, built to catch predictors that look right but aren't. Apache 2.0, 91 tests green.
View on GitHub →A sovereign Sanskrit retrieval engine — Pāṇinian grammar, three-channel φ-RRF fusion and MFDFA signal analysis. Scores nDCG@5 0.76, roughly twice the sentence-transformer baseline. Built falsifiability-first.
Private research · available on requestA structured-output prompting framework for LLMs, with an empirical study of non-monotonic format adherence across model scales — 4B, 8B and 70B.
Private research · available on requestFull-spectrum Sanskrit analysis — chandas (metre) detection, multilingual transliteration and semantic search. Held private until it meets the standard.
The unified fractal retrieval core beneath the Kaṇaja line — the heavy research engine, maturing toward release.
Latest — what I'm proving now
I extended the falsification gate from search to LLM free-text: a grounding filter that keeps only source-backed answers — and a gate that tests the AI judge itself, refusing any grader that can't tell a real answer from a fluent fake.
Writing
The discipline
"Truth alone triumphs — not falsehood."
I commit to never publishing a claim I don't believe will survive falsification. When my own results fail, I publish that too. The first work I shipped publicly retracted its own pre-fix numbers — and that retraction is what made the validated numbers trustworthy.
"You have the right to action only — never to its fruits."
I ship the work and let time decide its outcome. Whether it survives verification is the measure — not the forecast. Every artifact I publish has a thirty-second demonstration anyone can run.
Background
Undergraduate at RMIT University, Melbourne. I work on the small piece of statistical machinery that makes retrieval evaluation honest, and on cross-cultural Sanskrit–Dravidian retrieval — an under-served corner of computational philology whose hard parts push methodology in useful directions.
Open to collaboration on falsification methodology, low-resource philological corpora, or both. The brand behind the work is Bhardwaj & Sons — a house of standards.