SESSION 2026-09-14 · PLATES 14 · RENDER build/render.py · CHECK build/verify.py · JS 0
Most profiles are a showcase, where every row is a win. This one is a blotter, so it has a verdict column, and two of the four verdicts are not wins. That is the point. Hiding them would make the other two worth less.
| Work | Verdict | The record | |
|---|---|---|---|
01 |
A Regime-Aware Meta-Learning Framework for Selective Directional Trading in Cryptocurrency Markets | PEER-REVIEWED |
First author. Unsupervised temporal clustering finds latent market regimes; a MAML-inspired meta-learned classifier then abstains rather than guess under low confidence. IEEE ICIPTM 2026 · 10.1109/ICIPTM69057.2026.11466047 |
02 |
Regime-Route · C++20 PostgreSQL Redis Next.js |
FALSIFIED |
Proof-carrying execution: every routing decision emits a hash-verifiable receipt. Multi-tenant auth, idempotent submission, TLS reverse proxy. Replayed against 13M+ real order-book rows with paired counterfactuals. The honest, final conclusion was that no economically meaningful edge was found — 26 orders, −451.96 bps average edge, 38% win rate across $259,069,209 of paper notional. I published that instead of quietly reframing the goal. The receipts still work; so does the negative result. |
03 |
Tensor-Forge · C++20 WGSL Next.js |
SHIPPED |
A from-scratch JIT tensor compiler with no PyTorch and no CUDA underneath — it lowers and shape-specialises itself, and every stage is inspectable. 5/5 CTest suites passing, full CI. |
04 |
Bitcoin-Alpha-System · Python PyTorch |
UNDER AUDIT |
Under active audit and rebuild. Walk-forward and holdout validation are still in progress, so no performance number appears on this page and none should be stated as final. Honest status: in progress, rebuilding under audit. |
KEY — PEER-REVIEWED outside review passed · SHIPPED tested and running · FALSIFIED looked for the effect, did not find it, published anyway · UNDER AUDIT still validating, nothing claimed until it clears
A receipt is only worth something if a stranger can check it. Every routing decision in Regime-Route emits one, and anyone holding the same tape can recompute the digest and compare it byte-for-byte, no trust in my log required. **That machinery works.** Pointed at 13M+ real order-book rows with paired counterfactuals, it came back and said the effect is not there: the honest, final conclusion was that **no economically meaningful edge was found.** That is a feature, not a failure to hide, it is what validation discipline looks like when the answer is no.
raw tape ──> temporal clustering ──> regime label ──┬──> confident ────────> take the position
│
└──> not confident ──> stand down, stay flat
This is what the paper argues, drawn rather than described. Cluster the tape into latent regimes
with no labels, then let the classifier decline to act in the regime it cannot call. The lower
lane is the part that makes it a rule rather than a slogan: the model's own confidence, with the
line it has to clear before a position is allowed. A model that stands down 29% of the time and is
right when it speaks beats one that always has an opinion. The plate is a schematic of the rule,
not backtest output; the shape is illustrative, the argument is not.
No PyTorch and no CUDA underneath means the interesting claim is not that it runs; it is that
there is nothing below it doing the real work. A tensor graph is parsed, shape-specialised,
lowered, emitted as WGSL and dispatched, and you can read what came out of every one of those
five stages. 5/5 CTest suites, full CI. The plate is a schematic of the path a kernel takes, not
a map of the source tree.
COUNTED LIVE BY THE GITHUB SEARCH API, NOT BY ME
| Project | Maintained by | Where I worked | Merged |
|---|---|---|---|
| Gemini CLI | cli core extensions devtools |
8 |
|
| Oppia | Oppia Foundation | LEAP team — led a Redis infrastructure upgrade | 10 |
| OpenMetadata | Collate | metadata platform | 1 |
19 pull requests merged by maintainers who owe me nothing · counted on 2026-09-14
The Oppia one is the one I would point at. The Redis upgrade was unglamorous infrastructure work that was failing CI for every other contributor, which is exactly why it was worth doing.
Manav Sharma — Bachelors in Computer Science Engineering, Class of 2026. Looking for research and systems work where the validation is taken as seriously as the model. If a number on this page is wrong, open an issue: I would rather be corrected in public than quoted incorrectly.