Skip to content
View bishnubista's full-sized avatar

Highlights

  • Pro

Organizations

@secure-agentic-framework

Block or report bishnubista

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. Markdown is supported. 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
bishnubista/README.md

Bishnu Bista

Hands-on engineering leader building dependable AI-agent systems, evaluation infrastructure, and security tooling.

I turn ambiguous ideas into tested, operable products—from architecture and agent workflows to full-stack implementation, security boundaries, release engineering, and contributor experience. I maintain SAF-MCP, contribute to Promptfoo, and co-authored ConformalGuard, accepted at COLM 2026. Previously at American Express and Wells Fargo.

I am especially interested in startup conversations where technical leadership, product ownership, and hands-on founding-engineer execution matter.

At a Glance

  • Technical leadership: architecture, 0→1 product engineering, platform design, full-stack delivery, release operations, and open-source maintenance
  • AI systems: coding agents, LangGraph and multi-agent workflows, LLM evaluation and red teaming, human-approval boundaries, retrieval, and local-first AI
  • Real-time AI: voice pipelines with Pipecat, Deepgram, WebRTC/WebSockets, streaming interfaces, and structured function calling
  • Security: Model Context Protocol (MCP), threat modeling, detection engineering, least privilege, auditability, and privacy-preserving system design
  • Core stack: Python, TypeScript, LangGraph, React, FastAPI, Hono, Cloudflare Workers/D1/R2, SQLite, Swift/SwiftUI, and agent-development tooling
  • Based in: San Francisco Bay Area

Open-Source Leadership

SAF-MCP — Maintainer

A MITRE ATT&CK-style framework for documenting and mitigating threats in Model Context Protocol and agentic systems. My contributions span threat research, Sigma detection rules, security-control design, catalog validation, and contributor infrastructure.

Selected work:

Promptfoo — Contributor

Merged contributions to the open-source platform for testing and red-teaming prompts, agents, and RAG systems:

Additional upstream work includes a benchmark-backed audit-log read-replica fix in Pangolin, a Python library behavior fix in Supabase, and module caching in the American Express One App platform.


Research

Accepted at the Conference on Language Modeling (COLM) 2026.

ConformalGuard is a model-agnostic calibration layer for choosing pre-deployment action thresholds with finite-sample false-positive control under exchangeability. We evaluated it with fine-tuned BERT, Claude Sonnet 4, and GPT-4o on public email corpora, exposing the recall and robustness trade-offs of coarse zero-shot LM scores.

Code and reproducibility · Paper PDF · OpenReview record


Products and Developer Infrastructure

Project Technical focus
OpenBird Open-source, local-first personal AI memory for macOS. Text-first capture, transient on-device OCR, SQLite/FTS5/vector retrieval, cited answers, Ollama/BYO-model support, and bounded read-only MCP access.
ConfPilot Open-source conference program operations from CFP to published schedule. Role-scoped workflows, idempotent lifecycle transitions, React/Hono, and runtime ports for Cloudflare D1/R2 or Node/SQLite.
CyclePel Native SwiftUI indoor-cycling platform with FTMS smart-trainer control, structured ERG workouts, route simulation, and hardware safety gates. Currently private and being prepared for a future open-source release.
cc-statusline A focused Claude Code statusline for model, directory, Git branch, and output-style context.
PlanGate PLAN.md-driven development with explicit quality gates for Claude Code workflows.

Engineering Approach

  • Agent-native delivery: decompose work into bounded workstreams, use coding agents for implementation and review, and require exact verification before integration
  • Dependability by design: deterministic evaluations, human approval for consequential actions, audit trails, privacy boundaries, and explicit failure states
  • Full-lifecycle ownership: product definition, system architecture, implementation, testing, packaging, release operations, documentation, and contributor workflows
  • Leader who still builds: stay close to code and users while creating the technical systems that let a team move quickly without losing correctness

Community

  • Organize Bay Area events through The AI Agents Community
  • Teach workshops on Claude Code, Cursor, and AI-powered software development
  • Build and contribute in public across agent security, evaluations, local-first AI, and developer infrastructure

Connect

If you are building coding agents, agent infrastructure, AI security, evaluation systems, or ambitious developer products—and need a hands-on technical leader who can move from architecture to shipped systems—let's connect.

LinkedIn · X · AI events

Outside software, I run marathons and train for triathlons.

Pinned Loading

  1. secure-agentic-framework/saf-mcp secure-agentic-framework/saf-mcp Public

    SAF-MCP is a comprehensive security framework for documenting and mitigating threats in the AI Agent ecosystem.

    Python 361 94

  2. conformalguard-colm-2026 conformalguard-colm-2026 Public

    COLM 2026 accepted research artifact: conformal calibration and false-positive-controlled action gating for LM classifiers.

    Python

  3. openbird openbird Public

    Open-source, local-first personal AI memory for macOS with cited retrieval, Ollama/BYO-model, and read-only MCP access.

    Python

  4. confpilot-open-source confpilot-open-source Public

    Open-source, self-hosted conference operations from CFP to published schedule — React/Hono on Cloudflare or Node/SQLite.

    TypeScript

  5. cc-statusline cc-statusline Public

    Simple, clean Claude Code statusline displaying model, directory, git branch, and output style

    Shell 5

  6. plangate plangate Public

    PLAN.md-driven development with quality gates — a Claude Code plugin

    Shell 1