PranjalSailwal
I build products end to end: from the question worth asking to the thing that ships.
Heading toward technical product and programme work, plus the consulting and implementation roles next to it.
98.1%
of the cost increase happened after the click
11 of 11
planted changes caught in a live pass, no false positives
68
unit and integration tests
Audit me shows each figure's source, how to reproduce it and its seal check. Or .
CS graduate, 2026, Dehradun
Dehradun, India
30.32° N, 78.03° E
It starts with a number that looks wrong.
A campaign that spent far more and converted far worse. Sensors a few metres apart that disagreed. I can't leave a number like that alone, so the first thing I write down is the question.
19.2x
more spend in the larger campaign
73.8%
drop in post-click approved conversion
10 to 20%
gap between sensors placed side by side
Then: who is it hurting?
Students reading an air-quality dashboard that hides the disagreement. A growth team about to blame ad inventory for a conversion problem. A problem is only real once I know whose it is.
Decide what not to build.
Migration Guard could have produced a single AI “safety score”. I chose explicit, documented rules instead, so every PASS, WARN or BLOCK comes with a reason someone can check.
0
passwords, emails or environment variables collected
pre-alpha
the project's own status
Prove it with a number someone can check.
If I can't measure it, it doesn't go on the page. Where a figure is only my word, the page says so, and you can switch those off.
98.1%
of the cost increase happened after the click
11 of 11
planted changes caught in a live pass, no false positives
68
unit and integration tests
3 things I built
to find something out.
NitroDB shows how I go deep, AetherScan how I handle data I can't fully trust, and the growth lab how I turn a business question into an answer. Each one comes with its receipts. Every project also plays as a short film: pick one and watch it.
Keep scrolling: the page runs sideways.
NitroDB
“What actually keeps a database's data safe through a crash?”
An embedded LSM-tree storage engine, built from papers with no tutorial
68
unit and integration tests
42
concurrency outcomes checked, 0 failed
2.25M
point reads a second in a one-iteration smoke run
recorded real run 60,000 writes in a recorded run of the real engine, drawn as its own SSTable files on each level, from the first flush to after compaction
AetherScan
“How much can you trust a single air-quality reading?”
Air quality platform quantifying disagreement between 8 sensor networks
11
data sources, each with licence, rate limit and how provenance is kept
53
API route handlers
12
backend test files, each named test_*.py
recorded real data 366 days of Dehradun air quality from a file in my own files (PM2.5 from 5.4 to 155.3), drawn as a skyline that is tall in winter and low in the monsoon. The file states no source
Growth Intelligence Lab
“Why did the cost per conversion jump four times over?”
A growth analytics platform that finds why cost per conversion jumps 4x when spend rises 19x, using only public data
98.1%
of the cost increase happened after the click
19.2x
more spend in the larger campaign
73.8%
drop in post-click approved conversion
6 more, and every repo on GitHub.
The other screenings, the technologies behind all of them, and the smaller repositories I haven't written up yet.
See all the workEvery figure here shows its working.
I can prove some of these numbers, I only wrote some of them down, and a few are just my word. Each one carries a stamp, and you can press it to see where it came from. Or switch off everything I can't prove.
49 figures on this site, 2 of them only my word. Nothing is deleted: switch it back and they return.
The ledger
The figures that matter most, and how each one is known. Press a stamp for its receipt.
Where I'm headed: what to build, and why.
I'm a fresher, so I won't claim a title I haven't held. What I can show is how I work: the question first, then the people it's for, the decision and what I measured. That is the work I'm heading toward: technical product and programme roles, and the consulting and implementation work next to them.
Technical product work
I can sit with engineers because I build the things: every project here went from a question to something running, and the decision about what not to build is written down in each one.
0
passwords, emails or environment variables collected
Product analysis
Funnels, segments and significance tests on public data only. I traced a fourfold jump in cost per conversion to one lever and turned the finding into a ranked list of experiments.
98.1%
of the cost increase happened after the click
Planning, and writing it down
Scope, schedules, risk and agile delivery, studied properly, plus the habit of writing the reasoning down so it outlives the meeting.
19
of them in project management
I write things up.
Dev logs about why I made a choice, and four preprints. Writing it down is how I find out whether I understood it, and it is the part of a project people can still use after the code has changed.
AetherScan: trying to make sense of air quality data
Air-quality data exists everywhere, but it's scattered, inconsistent, and hard to read in one place. It's less about building a ‘feature-heavy’ app and more about building something that stays readable even as the data grows.
Read it on dev.to
SyncState: writing together, even when the internet isn't there
Collaborative editors assume stable connectivity. When that assumption breaks, the whole system falters — collaboration shouldn't stop just because the network does.
Read it on dev.to
AetherScan: A Quantum-Classical Hybrid Platform for Heterogeneous Air Quality Sensor Fusion and Real-Time Environmental Intelligence
The research write-up of AetherScan: fusing air-quality readings from heterogeneous sensor networks and showing how much they disagree instead of averaging the gap away.
Read it on ResearchGate preprint
Implementation and Performance Analysis of a Spectral Monte Carlo Path Tracer with Physical Optics Simulation and Adaptive Variance Reduction
A reproducible validation methodology for a physically-based renderer: BVH-accelerated ray traversal, Monte Carlo path integration, and spectral wavelength sampling, cross-validated against analytical ground truth at 1.81M samples/sec.
Read it on ResearchGate preprint
Geodesic Flow Computation on Parametric Surfaces via WebAssembly: A Differential Geometry Engine with Adaptive Numerical Integration
A WebAssembly engine that computes geodesic flow, Christoffel symbols and curvature on parametric surfaces in real time, with curvature-sensitive adaptive stepping.
Read it on ResearchGate preprint
Wavelength-Resolved Geodesic Propagation on Parametric Surfaces: Coupling Differential Geometry with Spectral Dispersion
Couples differential geometry with spectral dispersion — computing geodesic flow and curvature (Christoffel symbols, Gaussian and mean curvature) on parametric 3D surfaces with adaptive numerical integration.
Read it on ResearchGate preprint
Everything I've writtenThe log is still open.
This is read live from my GitHub account, private repos included (only the language and the time, never the names). A push I make shows up here within a minute.
Languages I have been writing in
- Python31%
- TypeScript29%
- JavaScript11%
- Go6%
- HTML5%
- TeX4%
Latest pushes
- 4mo agoGitlens
- 1mo agoNeurotract
- 1mo agoAetherScan
- 26d agoGrowth-Campaign-Funnel-Intelligence-Lab
- 18d agoappwrite-migration-guard
- 9m agoa private repo
27 public repos, 1 private. Checked just now.
If a number does not add up, write to me.
- seq 0001you arrived
- seq 0002write back