8000 vbsamuel (OceanBz) · GitHub
[go: up one dir, main page]

Skip to content
View vbsamuel's full-sized avatar
💭
Improve Lives & Create Joyful Experiences
💭
Improve Lives & Create Joyful Experiences
  • *** Pilgrim ***
  • USA

Block or report vbsamuel

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
vbsamuel/README.md

Ideas, Methods, & Paper Clips

Durable · Decomposable · Dynamic · Distributed Systems
Across software, machines, and silicon.

I work on systems where execution, state, resources, coordination, and failure behavior remain explicit under real constraints.


Selected Work

Agentic Systems Portfolio

Two executable agentic workflows: model-authored software defect repair and catalog-to-published-content reconciliation. The local vertical has demonstrated real DeepSeek inference, failing-before/passing-after Git correction with a persistent candidate commit, and authenticated HTTP catalog updates with independent readback. A subsequent two-model run generated a policy-grounded catalog correction and stopped for explicit approval; model-driven publication and production qualification remain open.

repository · execution architecture · observed qualification evidence

Inference systems experiments

Controlled experiments in inference serving and accelerator behavior. Each experiment keeps the failure case, the recorded measurements, and the boundary of what the data can support.

Experiment Recorded result What changed
KV cache tiering warm TTFT 9.315 s → 0.336 s reload wins only after reuse amortizes persistence and transfer; cold path regressed
Dynamic batching 1,438.9 → 2,923.3 QPS at the best measured configuration oversized configured batches added queueing overhead when offered concurrency could not fill them
Precision × batch 3,271.6 → 44,494.9 samples/s across the measured sweep throughput was still rising at the largest tested batch; no saturation knee claimed
Baseline drift apparent 5.77% gain became 0.93% against the stabilized baseline candidate rejected as INSUFFICIENT_EVIDENCE

repository · evidence manifest

ONTEXA

A local-first biomedical discovery system built as a set of explicit services rather than a model wrapper: dataset search, graph retrieval, ontology processing, ranking, local model orchestration, metrics, and an operator surface.

scientific sources
      ↓
ingest / ontology
      ↓
graph + search
      ↓
ranking
      ↓
local model orchestration
      ↓
operator surface

The implementation spans Go dataset search, Rust ranking, Python orchestration, Neo4j graph state, PostgreSQL, and a React operator surface.

repository · orchestration · graph schema · ontology preprocessing

Public work appears here only when the implementation or evidence can be inspected directly. Forks, references, coursework, and unfinished prototypes are not part of this index.


Current Builds

These names are reserved for systems being developed as independent, testable implementations. They are not listed as completed work until code, qualification, and measured evidence are public.

Project Engineering problem
NodeFit determine whether a workload can safely execute on a measured machine and where its binding resource constraint lies
TypeMap identify where data loses meaning across type systems, formats, storage engines, and execution boundaries
SafePath verify whether consequential operations remain correct across effect, persistence, acknowledgement, and retry boundaries
Rebound enumerate and qualify the states reachable after interruption, restart, reconciliation, and recovery
SourceIQ reconstruct which prior execution, state transition, or external effect produced an observed result
DataHop preserve typed object identity and ownership while data moves across runtimes and memory domains
RunPack capture the inputs, state, effects, timing, dependencies, and evidence required to inspect and reproduce an execution

System Properties

Durable — state, identity, provenance, recovery, and semantics that survive process, machine, and time boundaries.
Decomposable — explicit components, interfaces, ownership, state machines, and independently testable behavior.
Dynamic — execution that responds to changing workload, topology, resources, latency, and operating conditions.
Distributed — computation, state, authority, and coordination across processes, machines, accelerators, and people.


Working Areas

Area Current questions
Execution scheduling, placement, batching, streaming, backpressure, retry, checkpointing, recovery
State ownership, journals, memory, causality, reconciliation, durable transitions
Coordination routing, delegation, attribution, synchronization, machine-to-machine and agent-to-agent interaction
Compute CPU/GPU execution, heterogeneous resources, locality, allocation, runtime control
Representation intermediate representations, structured content, spatial systems, progressive rendering, interaction surfaces
Verification invariants, qualification, falsification, fault injection, reproducibility, measured evidence

Engineering Rules

state            → one explicit authority
work             → bounded by real resources
mutation         → declared path
failure          → defined outcome
retry            → idempotent effect
coordination     → retained causality
measurement      → workload + environment + counter-result
claim            → code, test, measurement, or reproducible evidence

Repository Index

Systems — substantial original implementations
Components — reusable subsystems
Experiments — bounded technical hypotheses
Research — technical investigations and implementations
References / Forks — upstream work retained for study
Archive — historical work not representative of current engineering

The public repository count is not the portfolio. The index above is.


Earlier Scale

Meta      XR input + developer/platform programs        100+ B2B customers
PayPal    safety / security / risk platforms             300M+ users · $500B+ TPV
          engineering + operations                       500+ org

eBay      separation / personalization / operations      large-scale platform change

Work in my scope included reducing reporting latency from roughly 24 hours to under a minute, reducing cart abandonment by about 10%, and scaling PayPal Credit from launch to $1B+ ARR.

LinkedIn

Popular repositories Loading

  1. Census_Tracts Census_Tracts Public

    US Population Data File input is processed to provide an output file for relevant data for insights. The work involves data cleaning, standardization, and derivation of insights for further analysis.

    Python

  2. Sentiment-Analysis Sentiment-Analysis Public

    This Project provides a mechanism to injest social media content and perform deeper analysis on the part-of-words to determine the sentiment of language. The Sentiment of the language is scored bas…

  3. vbsamuel vbsamuel Public

    Config files for my GitHub profile.

    < 46BB /div>
  4. OpenDevin OpenDevin Public

    Forked from OpenHands/OpenHands

    🐚 OpenDevin: Code Less, Make More

    Python

  5. ai-agents ai-agents Public

    Forked from viktoriasemaan/ai-agents

    AI Agents with Bedrock

    Python

  6. MCP-servers MCP-servers Public

    Forked from modelcontextprotocol/servers

    Model Context Protocol Servers

    JavaScript

0