Prysai Lab · Contributions · Core projects · Star history
I build and inspect AI systems that hold up beyond the demo: agents, retrieval, evaluation, and tool use in real developer workflows.
My background includes six years in network operations and four years as an independent developer, overlapping within the same broader period. This combination informs my current scope across operational continuity, technical implementation, and delivery coordination.
Since 2024, I have also been developing full-quantization training and fine-tuning workflows for image AI models.
AI agents · retrieval · evaluation · developer tools · reliable delivery
I work through the full agent engineering loop: choose the right model and context, connect tools safely, measure quality and failure modes, and turn what works into repeatable workflows.
The standard is simple: fewer impressive demos, more systems that can be tested, understood, and improved.
Prysai is my studio for product experiments and practical work around AI systems and developer tools—small, testable pieces designed to be inspected, measured, and reused.
|
FLAGSHIP · EVIDENCE-LED LEARNING SYSTEM Prysai LLM Playbook A public, six-language learning and practice system for reliable work with LLMs, Codex, skills, tools, and agent workflows. It pairs concepts with observable labs, acceptance checks, and source boundaries. repository · documentation · v0.1.0-alpha |
Public development candidate: the structure and checks are available to inspect; broader learner validation and independent review remain in progress.
A selective record of changes that made it into other people's codebases, docs, and releases. Each entry links directly to the merged work.
| Proof of work |
|---|
| COLLABORATION & TEST RELIABILITY jupyterlab/jupyter-chat · PR #486 · merged Eliminated SQLite YStore lock contention in parallel UI tests by giving each test server an isolated YStore database. |
| AGENT RUNTIME & OPERATIONS microsoft/agent-framework · PR #7622 · #7597 · merged Made rejected MCP skill archives visible at the operational warning level and preserved Mistral prompt-cache usage details across regular and streaming responses. |
| AGENT CLI & INTEGRATION gnt-ai/gnt · PR #178 · #177 · #176 · #175 · #174 · #173 · #172 · #168 · #164 · #163 · #162 · #138 · merged Expanded CLI automation with JSON output for rule linting, organization inspection, and key listing; added connector coverage for Figma, GitLab Threads, and HubSpot; and strengthened terminal formatting, repository conventions, documentation, and review output. |
| ROUTING & OBSERVABILITY vllm-project/semantic-router · PR #2907 · #2896 · #2874 · merged Corrected Router Replay attribution, wired configured Milvus consistency through cache reads and collection setup, and removed a racy Dashboard assertion while retaining stable lifecycle guarantees. |
| VOICE AI & CONFIGURATION livekit/agents · PR #6810 · merged Added explicit Google Speech-to-Text project selection while preserving credential-based discovery for existing callers. |
| CONNECTORS & PRODUCT INTEGRATION oomol-lab/open-connector · PR #332 · #300 · merged Expanded the Resend connector from basic sending into batch, scheduling, retrieval, cancellation, and attachment workflows, then made run-log summaries resilient to malformed URL-like values while preserving redaction and valid sibling fields. |
| VOICE AI & AUDIO PIPELINES pipecat-ai/pipecat · PR #5298 · merged Made AssemblyAI speech recognition inherit the pipeline input sample rate unless callers provide an explicit override. |
| EVALUATION & OBSERVABILITY thangldw/ragops · PR #30 · merged Added an offline Phoenix recorded-score adapter with deterministic case-ID joins, synthetic fixtures, documentation, and tests. |
| AGENT TOOLING & COMPATIBILITY ag2ai/ag2 · PR #3148 · merged Added HEIC and HEIF MIME types to the Gemini image-input type path and covered the public factory with regression tests. |
| AGENT TOOLING & COMPATIBILITY mlflow/mlflow · PR #25010 · #25000 · merged Preserved provider-specific usage metadata across Gemini, Anthropic, and OpenAI-compatible gateway adapters, with regression coverage for normalized and nested usage details. |
| AGENT TOOLING & DOCUMENTATION jupyterlab/jupyter-ai · PR #1639 · merged Added a concise interaction model explaining how chat, personas, ACP agents, model providers, and MCP tools fit together. |
| MEMORY & OPERATIONS MSKazemi/kubeintellect · PR #116 · #111 · #105 · merged Made the memory-recall similarity floor configurable, updated the kube-q Homebrew formula, and documented v4 data-handling boundaries. |
| LANGUAGE ANALYSIS & RELIABILITY squid-protocol/gitgalaxy · PR #1761 · #1724 · #1710 · #1657 · #1655 · #1654 · #1639 · #1201 · merged Improved detector correctness across TypeScript, C#, and C++: recovered missed brace-less functions, eliminated phantom matches, recognized readonly/ref structs, avoided digit-separator misclassification, and prevented silent archetype-metric dimension truncation. |
| AGENT TOOLING & COMPATIBILITY mastra-ai/mastra · PR #21559 · #21281 · #21147 · merged Kept array text-streams valid for batched first chunks, resolved dynamic models once per assigned-tool batch, and fixed CommonJS schema compatibility by bundling the Zod conversion dependencies. |
| RELIABILITY & EXPERIENCE felladrin/MiniSearch · PR #2390 · #2387 · #2386 · #2385 · #2377 · #2375 · #2372 · #2370 · #2367 · #2339 · #2324 · merged Reworked the search experience with a proper landing page, clearer result hierarchy, and reduced-motion-aware drawers; hardened it end-to-end for IME and non-Latin follow-ups, failure states, early rate limiting, SSRF-safe capped thumbnail fetches, and robust error handling. |
| AGENT ORCHESTRATION & CLI a5c-ai/babysitter · PR #1727 · merged Made process-backed run creation safely adopt the active session-start bare run while retaining explicit-run and cross-boundary create-new safeguards. |
| PRIVACY & EVALUATION seva9523/EvalRepro · PR #12 · merged Added --no-id_preview to suppress sample ID previews while preserving sample and field digests. |
| DOCUMENTATION & DEVELOPER EXPERIENCE abenneto/gandharva · PR #19 · merged Refined README formatting and section headers. |
Every entry above links to a public pull request confirmed as merged into its upstream repository.
A focused working set of upstream projects I study, use, or contribute to. The selection is about technical leverage, not repository size.
|
RAGOps Evaluation infrastructure Offline evaluation and explainable release gates for RAG systems and agents. upstream · my fork |
PydanticAI Typed agent runtime Typed tools, structured outputs, and validation for production-minded agents. upstream · my fork |
|
FastMCP Tool connectivity Protocol infrastructure for connecting MCP servers and clients. upstream · my fork |
Promptfoo Quality and safety gates Evaluation and red-teaming for prompts, agents, and RAG pipelines. upstream · my fork |
|
Browser Use Real-world execution Browser control for tool-using agents and workflow automation. upstream · my fork |
Spec Kitty Developer workflow Spec-driven development and repeatable worktree workflows for serious projects. upstream · my fork |
These links point to upstream projects and my forks. They are study and contribution paths, not claims of original ownership.
One account-level signal: the total stars received by all public repositories over time. It is a trend for the account, not the star count of any single project.
All public repositories are included · forks contribute only their own stars · refreshed daily by GitHub Actions.
AI agents · RAG · LLM evaluation · MCP · developer tools · Codex workflows
Building in public, one reliable workflow at a time.