Krishnendu Biswas

Work

AI agent systems and machine learning projects. Running since Sep 2026 on open-source Hermes Agent with human approval gates. Machine learning projects focus on NLP fine-tuning, class imbalance, and analytics dashboards.

The platform runs 57 recurring scheduled jobs across 9 agents: 8 that each own one area, plus 1 that routes my requests. Operations like sending emails, deleting data, or public git pushes pause at approval gates until I verify them.

Agent systems (7 case studies)

  • Multi-agent operations platform

    Nine agents that hand work to each other on task cards, with a review step before anything risky. Includes hourly drift monitoring and janitor cleanups under infrastructure.

    9 AGENTS · 57 JOBS

  • Model routing and tiered compute

    One gateway on my server in front of most of my model accounts, so a failed call moves to a backup model. Claude models are called directly. Plain Python and Bash scripts perform exact calculations before any model is invoked.

    GATEWAY & FALLBACK

  • Fast decisions

    A small typed decision model (System 1) that answers narrow questions in ~300 ms with a choice or score, keeping large chat models quiet for simple triage.

    TYPED DECISIONS

  • How my agents remember: five layers of memory I can read and edit

    Each agent keeps short curated notes, and a context engine (LCM-X) summarises long chats into searchable summaries while keeping the full original messages on disk.

    CHAT MEMORY

  • Reliability and security

    Encrypted nightly backups copied offsite to cloud storage and a private GitHub repository, locked-down secrets, and a written, scripted runbook to rebuild the server from scratch.

    ENCRYPTED BACKUPS

  • Finance ledger agent

    Reads my receipt emails (read-only access), turns them into ledger entries with fixed parsing rules, and nets refunds against the original order in my self-hosted Firefly III finance app.

    RULE-BASED PARSING

  • Personal LinkedIn studio

    Local drafting environment that screens technical posts against voice notes and anti-AI writing patterns. Every post is manually reviewed and published by me.

    HUMAN PUBLISHES

Private operational platform. Architecture, system boundaries, and verified telemetry are presented across each case study.

Data: Token ledger, 12.47 B TOKENS across 3 sources

My machine learning projects, four with full write ups, are on the projects page