Skip to content
View Kesh3805's full-sized avatar
💗
Hehee
💗
Hehee

Highlights

  • Pro

Block or report Kesh3805

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Kesh3805/README.md

I build computational systems and investigate how their behaviour changes under constraints — through implementation, simulation, and controlled experiments.


Work

temporal-compute · Rust, Python · Research

A deterministic simulator for computation whose value depends on when it finishes. The question it investigates: if a correct answer arrives after the decision it was meant to inform, was the computation successful?

Experiment TC-S1 tested two temporal-utility schedulers against four conventional baselines across 200 synthetic scenarios, with the protocol preregistered before the run.

TC-S1: assumed value curves and measured totals
Left: utility against completion time for six real generated tasks from one TC-S1 scenario. Right: measured totals for six scheduling policies over 200 scenarios.

The primary scheduler scored 524,509 against EDF's 552,525 — a 5.07% regression — and failed all five preregistered gates. A secondary variant came within 0.35% of EDF. A second protocol, allocating render samples by marginal value, is registered and has not been run.

A decision event with its past light cone shaded; a computation that completes after the decision, outside its past cone.

reviewgraph · Rust, TypeScript · In progress

ReviewGraph investigates whether pull-request findings can be backed by explicit dependency and change-impact evidence from a code graph, rather than relying entirely on textual diffs.

Built so far: TypeScript parsing into a persistent code graph, change-impact analysis across both base and head graph states, risk scoring, and reviewer prompts. The verification stage stores and deduplicates findings; the checks that would re-test a finding against the base graph are not written, and it has not reviewed a real pull request end to end.

Impact graph from the test fixture, and pipeline status
Impact graph from a golden test fixture (synthetic pull request) with six impacted elements, and the seven pipeline stages marked built, partly built, or not started.

Layered-Memory-Architecture · Python · Experimental

A self-hosted conversational system for testing retrieval behaviour across a tiered memory store.

The most significant result is a regression: on a 605-chunk corpus, adding BM25 to vector search lowered MRR from 0.758 to 0.700 over 30 queries. A separate configuration — same corpus, same query set — added a reranker instead and scored 0.772. The hybrid retrieval regression is the unexplained result; the reranker number is a subsequent data point, not a resolution.

MRR result and memory tier structure
Mean reciprocal rank over 30 queries: vector search 0.758, with BM25 hybrid 0.700, with reranker 0.772. And the four memory tiers.

Research interests

Mathematical structures have a habit of describing physical reality more precisely than intuition suggests they should. The Schwarzschild metric predicts signal behaviour near a black-hole horizon with enough specificity to draw the diagram below — computed from the equations, not by hand.

Penrose diagram of a Schwarzschild black hole, computed
Penrose diagram of a Schwarzschild black hole. An observer falls inward emitting 16 light signals at equal proper-time intervals; a static receiver at r = 3.6M receives them progressively later. Signals emitted strictly inside r = 2M end at the singularity; one emitted exactly at r = 2M stays on the horizon. Neither reach the receiver.

Schwarzschild metric (G = c = 1), Kruskal–Szekeres coordinates compactified with arctan. Thick curve: observer falling from rest at r = 2.5M. Thin curve: receiver at r = 3.6M. Ink signals reach the receiver; orange ones, emitted strictly inside r = 2M (U = const > 0), end at the singularity. Full notes: DIAGRAM-NOTES.md.

Whether mathematical structures like this are discovered or constructed is an open question. What it means for computation to be bounded by physical time is a more practical one — it is what temporal-compute is built around.

Pinned Loading

  1. Layered-Memory-Architecture Layered-Memory-Architecture Public

    Self-hosted, policy-driven RAG chat system with intent classification, FastAPI, PostgreSQL + pgvector, and React observability UI.

    Python 1

  2. reviewgraph reviewgraph Public

    Graph-native, incremental, evidence-driven pull request reviewer using persistent CodeGraph analysis.

    Rust

  3. temporal-compute temporal-compute Public

    A deterministic runtime and research platform for scheduling computation by temporal utility, freshness, and deadlines.

    Python

  4. ACFS ACFS Public

    AI Content Forensics Suite: Multimodal AI content detection (ViT, DeiT, RoBERTa) with explainable forensic analysis.

    Python