Skip to content
View harshkumawat5's full-sized avatar
🎯
Focusing
🎯
Focusing
6 followers · 0 following

Achievements

Achievement: Pull SharkAchievement: Quickdraw

Achievements

Achievement: Pull SharkAchievement: Quickdraw

Block or report harshkumawat5

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
harshkumawat5/README.md 
class Harsh:
    role      = "Backend Engineer @ Irdeto"          # OTT & PayTV platforms
    based_in  = "Jaipur, India 🇮🇳"
    day_job   = ["Java", "Spring Boot", "gRPC", "Python", "AWS", "BPMN"]
    side_quest = "evals & guardrails for AI agents"
    shipped   = {
        "ingestion throughput (media)":     "3.5x  (1K → 3.5K/day)",
        "ingestion throughput (non-media)": "1.9x  (4K → 7.5K/day)",
        "DSA problems solved":              "1500+",
    }
    belief    = "Don't trust the agent's 'Done!'. Diff the final state."

🤖 What I'm building: agent evals

📬 said-vs-did

When an inbox agent says "Done, I've declined the 3pm", did it? Mock Gmail + Calendar, 25 tasks, graded from final mailbox state, every claim fact-checked.

→ 225 runs · Haiku / Sonnet / Opus · 0 false claims, and one "it failed" that had actually succeeded.

🛡️ salus-shadow-hook

A shadow-mode Claude Code PreToolUse hook with 5 safety policies, replayed over 51 real SWE-agent trajectories.

→ 1,629 calls · caught an rm -f *.py that deleted matplotlib's setup.py, and SWE-bench still scored the task as resolved.

🏁 superset-race

Give coding agents the same prompt in parallel git worktrees, run your tests in each one, get a scoreboard.

→ Claude Code vs Codex on a real python-humanize bug: 770 ✅ vs 676 ✅ / 90 ❌.
🧪 More projects
Project What
🌍 latency-topology-visualizer 3D globe of crypto exchange and cloud-region latency · Next.js + Three.js
🏗️ Custom_Object_detection YOLOv4/v7/v8 + TF-OD for symbols in architectural floor plans · University of Essex research internship
📚 machine_learning_revision Core ML algorithms, end to end in notebooks

🛠️ Tech stack

📈 Activity

contribution snake

Pinned Loading

  1. machine_learning_revision machine_learning_revision Public

    Notebooks covering core ML algorithms end to end

    Jupyter Notebook 3 1

  2. Custom_Object_detection Custom_Object_detection Public

    Symbol detection in architectural floor plans with YOLOv4/v7/v8 and the TF Object Detection API (University of Essex research internship)

    Jupyter Notebook

  3. latency-topology-visualizer latency-topology-visualizer Public

    3D globe of crypto exchange and cloud-region latency, with live and historical views (Next.js + Three.js)

    TypeScript

  4. said-vs-did said-vs-did Public

    Does the inbox agent's 'Done!' match the mailbox? 25 email/calendar tasks graded from final state, every claim checked.

    Python

  5. salus-shadow-hook salus-shadow-hook Public

    Shadow-mode Claude Code PreToolUse hook (5 safety policies) replayed over 51 public SWE-agent trajectories

    Python

  6. superset-race superset-race Public

    Race coding agents on one prompt in Superset workspaces, verify every worktree, print a scoreboard

    TypeScript