Mukunda Rao Katta

Karna — Your Loyal AI Agent Platform. Self-hosted personal AI assistant with multi-channel messaging, extensible skills, and semantic memory.

Rudra Narayana Hindu Temple — community platform for events, donations, and priest scheduling

Standalone AI agent runtime — tool execution, context management, and multi-model routing for autonomous assistants

Chetana — AI Consciousness Research Platform. Test AI models against 14 consciousness indicators from 6 scientific theories.

Pull-model episodic memory plugin for Hermes Agent. Real deletes, audit trace, BYO Claude. MIT.

BuildCost — Construction Cost Estimator. AI-powered construction cost estimation from blueprints.

Linter for MCP (Model Context Protocol) config files. Python, CLI + library. Supports Claude Desktop, Cursor, Cline, Windsurf, Zed.

Crack the entrance exam. — AI-tutored prep for JEE, NEET, MCAT, CAT, and the bar. One subject at a time, one question at a time

Aphrodite — AI UI Component Library. AI-generated UI component library

Demeter — AI-Native CMS. AI-native content management system

Design your tattoo. Before the needle. AI tattoo designer with flash from actual shops. See it on your skin before you commit.

Swarm intelligence multi-agent framework — orchestrate collaborative AI agent swarms with voting and consensus

Personal portfolio — mukundakatta.dev | AI/ML Engineer

Public hub for my AI SDK, MCP, eval, and developer tooling contributions

Profile README — AI/ML engineer portfolio, open-source contributions, and featured projects

Agentic content studio: one product → 5 platform-tailored drafts → posted from your own browser session. 15 channels, no third-party API keys.

AI art community platform built with React Native, Expo, and Supabase

Sadhak (साधक) — AI-powered job search pipeline built on Claude Code. The Seeker.

Wire-level prompt introspection for LLM SDK calls. See exactly what was sent, with credentials redacted by default.

Small, opinionated AWS Bedrock client wrapper: adaptive throttle, cache-aware cost tracking, structured-output parse-and-repair. Single-cloud, single-purpose.

A curated list of tools, papers, datasets, and guardrails for defending LLMs against prompt injection.

Real-time school bus tracking with AI arrival prediction

Clone yourself on video. — Record once. Speak any script, in any language, on every platform. Made for creators, priced for cre

Your inbox, on autopilot. — AI reads every email, sorts it, summarizes threads, and drafts replies in your voice. All for five d

Therapy. Built for teenagers. A text-first mental health app that actually talks like a teen — and escalates to a real human when

Thumbnails that actually get clicked. Describe your video. Get ten MrBeast-quality thumbnails in thirty seconds, every time.

Read contracts like a lawyer. Paste a contract. Get the red flags, the boilerplate, and the terms worth negotiating. In plain Engl

Send money home. Instantly. Cheaply. — Stablecoin rails from the US and GCC to India and the Philippines. Settles in seconds, not days.

Your all-nighter research assistant. Pick a topic. Scholar reads fifty sources, synthesizes the arguments, and hands you an outline with

Therapy that gets it. — AI-first emotional support for people who share your context. New dads. Immigrants. First-time found

Know your rights. In plain English. — AI legal copilot for landlords, tenants, freelancers, and small claims. Cheaper than a lawyer. Faste

Sing it in anyone's voice. Covers, duets, parodies. Generate studio-quality vocals from your favorite artists — ethically licen

The sleep coach your watch forgot. — Plug in your Oura, Whoop, or Apple Watch. Get an AI coach that tells you what to actually do differe

Every Google review, answered overnight. We draft a response to every review — 5 star and 1 star — in your voice. You approve in bulk each mo

The resale app for kids stuff. — Buy and sell outgrown clothes, toys, and gear. Local pickup. Fixed prices. No haggling, no scams.

The reading coach for your first-grader. Kids read aloud. ReadMate listens, gently corrects, and celebrates. Parents get a progress report ev

Contractors — stop writing quotes by hand. Describe the job in plain words. Get an itemized quote with labor, materials, and markup, ready to s

Your newsletter. Ghostwritten. Tell us what happened this week. Wake up tomorrow to a full issue, in your voice, ready to send.

Nail your next interview. — AI-powered mock interviews that feel real. Role-specific questions. Feedback that actually helps you

Cursor, but for product managers. — Synthesize customer calls. Draft PRDs. Ship roadmaps. The AI workbench PMs have been waiting for.

Your social media team, in a box. — Photos in. Posts out. AI writes the captions, picks the hashtags, and schedules it all for you.

Dub your videos. In every language. — Upload once. Get perfectly lip-synced dubs in 40 plus languages, at creator-friendly prices.

Your dietician, in your camera roll. Snap your plate. See macros, calories, and nutrient gaps. Get a friendlier suggestion for tomorrow.

Your menu. Digitized. Nutrition-calculated. Photograph your handwritten menu. Get a clean digital version with calories, allergens, and QR-ready

Agency proposals, written in ten minutes. Tell us about the client and the scope. Get a polished proposal with pricing and timeline, on your l

The editor for indie podcasters. Drop your raw recording. Get a clean cut with chapters, show notes, and a transcript — ready to publ

The code review bot that actually helps. Reviews every PR in minutes. Flags the bugs, suggests the fix, and explains why — like a senior engi

Your cycle. Understood. A cycle tracker that respects your privacy and actually gives useful coaching — for energy, sleep, a

Bring old photos back to life. — Restore grainy family photos in one tap. Turn them into short videos you actually want to share.

Math that clicks. One AI tutor per grade, K through 10. Homework help. Test prep. Gentle drills. All in one app.

Studio-grade headshots. From your selfies. Forty professional headshots in twenty minutes. No photographer, no studio, no suit required.

An Instagram grid that actually looks good. Plan six months of posts visually. See your grid before you post it. Never break the aesthetic again

Taxes for the 1099 life. — Track income, expenses, and estimated taxes for rideshare, freelance, and every side hustle in betwe

What's for dinner? Ask your fridge. Open the door. Snap a photo. Get three dinner ideas using only what's inside. No grocery trip requir

Meetings that actually ship. — An AI teammate that joins every call, files Jira tickets, Slacks the follow-ups, and books the next

One podcast. Sixty shorts. Drop your hour-long podcast in. Get sixty scroll-stopping vertical clips with captions, tuned per pl

The CRM built for a team of one. Freelancers, consultants, solo founders. Track clients, deals, and follow-ups without Salesforce-lev

Coding for teens. That doesn't feel like school. Your kid builds a game, a Discord bot, or a mod — guided by an AI tutor that speaks their language.

Your AI scribe for the exam room. Listens to the patient conversation. Writes the SOAP note. Cuts physicians' documentation time by 75

Your paid community, done right. — A clean, fast home for subscribers. Messaging, events, gated content. No Discord chaos.

Cancel every subscription you forgot. Connect your card. See the 23 things you're paying for. Cancel any in one tap. Keep what you love.

Five minutes. Regulate your nervous system. Guided breathwork for anxiety, sleep, and focus. No subscription paywall on the first week.

The receipts. All of them. Sorted. Email them, photograph them, forward them. Boxit files each one under the right category — ready for

One form. Thousand applications. — Fill out your profile once. Our AI finds and applies to every job that fits, with tailored cover let

The AI receptionist that never sleeps. — Dentists, salons, law firms, trades. An AI that answers every call, books appointments, and sends re

Turn any photo into anime. One tap. Six seconds. Ready to post. Studios Ghibli, Shinkai, shonen — pick your vibe.

Your dog. As Napoleon. AI portraits of your pet in any style — royal, Renaissance, astronaut, wizard. Made for your wall.

Stop undercharging. Freelancers, we see you. Tell us the project, the city, the seniority. We tell you what to charge — backed by real industry r

Strip Markdown formatting (headers, bold, italic, links, code, blockquotes) to plain text. Conservative, fast, zero deps.

No description.

No description.

Health status tracker for LLM agent components

In-memory LRU cache for LLM responses keyed by request hash

Deterministic mock LLM for testing agent pipelines

No description.

No description.

Mask known secret patterns (API keys, JWTs, AWS access keys, GitHub tokens) in log lines before they reach stdout/files/sinks. Zero deps.

Parse 429 / rate-limit responses from Anthropic, OpenAI, Google Gemini, and AWS Bedrock into a unified shape: retry_after, kind (RPM/TPM/concurrent), provider.

Landing repo for the bedrockcache + bedrockstack + ragvitals Python libraries: production-grade AWS Bedrock + Anthropic Claude.

Drop-in safety layer for Hermes Agent: arg validation + egress allowlist + budget cap + structured output. Built for the Hermes Agent Challenge on dev.to.

Python port of agentidemp-rs: idempotency key helpers for LLM and agent calls. sha256-hex, UUIDv5, scoped helpers.

Python port of tool-output-truncate: truncate LLM tool output with head/tail/middle/middle_lines strategies. UTF-8 safe, zero deps.

Pre-warm Anthropic prompt cache before user traffic. Pairs with cachebench.

Multi-provider failover for LLM calls. Anthropic to OpenAI to Gemini on retryable errors.

Thin Anthropic Message Batches helper. Submit, poll, retrieve, with cost tally.

Detect when an LLM agent gets stuck calling the same tool with the same args. Sliding-window loop detector. Zero deps.

Semver-pin LLM system prompts. Register name+version+content, retrieve by pin, persist to JSON. Zero deps.

Cooperative per-task deadline primitive for agent workflows. check_or_raise, remaining_seconds, intersect for nested ops. Zero deps.

Approximate token counts for LLM messages, system prompts, and tools. chars/4 by default, BYO tokenizer callable. Zero deps.

Content-addressable LRU cache for LLM agent tool calls. Hash(tool_name, args) → cached result. TTL + LRU eviction. Zero deps.

Generate Anthropic / OpenAI tool schemas from a Python function signature plus a Google-style docstring. Zero deps.

One registry per LLM agent tool: pair a callable with its schema, side-effect tags, and default args. Dispatch by name; build Anthropic/OpenAI tools list. Zero deps.

Sync and async timeout enforcement for LLM agent tool functions — zero dependencies

No description.

Validate LLM conversation messages before sending — Anthropic and OpenAI rules, zero dependencies

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

Tag and filter agent runs with arbitrary string labels

No description.

Extract usage, model, and stop-reason metadata from LLM API responses

Sanitize user input before sending to an LLM

Per-turn conversation statistics for LLM applications

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

Export LLM conversations to markdown, HTML, plain text, and JSON

Registry of named agent capabilities with enable/disable, tags, and prerequisites

No description.

Composable system prompt builder for LLM agents

Lightweight structured plan with step status tracking for agent runs

No description.

No description.

No description.

No description.

Python port of @mukundakatta/pii-sentry: detect and redact PII and secret-like values.

Record consent-aware redactions for privacy review trails. Python port of @mukundakatta/consent-redaction-log.

Generate balanced eval cases from bugs, docs, examples, and policies. Python port of @mukundakatta/eval-dataset-smith.

Python port of @mukundakatta/vector-poison-score: detects vector-text mismatch, zero/NaN vectors, instruction-like payloads, link farms.

Threat-scoring library for AI-app security monitoring. Python port of @mukundakatta/kavach.

Wire-level prompt introspection for LLM SDK calls. See exactly what was sent, with credentials redacted by default. Anthropic, OpenAI, any httpx-based client.

Prompt-cache observability for LLM APIs. Per-call hit ratio, cost saved, regression alerts. Anthropic, OpenAI, Bedrock.

Aggregate LLM API cost across providers, models, and time windows. Provider-agnostic. Zero deps.

Fleet-level batch dispatcher for LLM APIs. Pool requests across tasks, route to provider Batch APIs, save 50% on cost without rewriting your agent loops.

Classify LLM provider error responses (rate-limit, auth, server, context-window, content-policy). Anthropic, OpenAI, Google, AWS Bedrock. Zero deps.

Shared token + dollar budget across concurrent LLM tasks. Thread-safe, zero deps.

Idempotency keys for LLM agent retries. Deterministic content-derived keys (UUIDv5 or sha256-hex) so retries dedupe at the provider.

LRU cache where eviction is weighted by token count, not entry count. Bound a prompt cache by tokens (or any other size unit) instead of N entries. Zero deps.

Deterministic cache key for an LLM prompt: normalize whitespace, hash messages, mix in model + temperature. Pairs with semantic-cache-key. Zero deps.

Thread-safe shared token + USD budget for concurrent LLM tasks. Reserve/commit two-phase API for fan-out workloads. Sibling to the Rust crate token-budget-pool. Zero deps.

Deterministic cache key for an embedding request: hash text + mix in provider, model, and dimensionality. So a cache survives model upgrades without false hits. Zero deps.

Phoenix-compatible exporter: ragdrift's five drift scalars as OpenTelemetry span attributes for Arize alerting.

Regex-only PII detector for emails, phones, SSNs, credit cards, and prefixed API keys. Rust port of pii-sentry. Zero deps.

Drop-in USD caps + egress allowlist + audit log for the Hermes Agent runtime.

Stable IDs for agent steps: deterministic hash of (run_id, step_index, kind) so events and traces share keys across reruns. Zero deps.

Tiny provider failover chain for LLM calls. BYO provider callables, async + sync, custom skip predicate, attempt trace.

Python port of llm-message-hash: canonical sha256 hash of LLM request structures. Per-provider presets drop noise fields. For cache keys and idempotency.

Zero-dep PII and secret scrubber for AI agent logs, traces, and audit output.

In-process LRU cache for LLM responses. TTL, sync + async, decorator + direct API, thread-safe. Zero deps.

Classify exceptions from LLM agent tools into a closed ErrorKind enum (USER_INPUT/AUTH/RATE_LIMITED/...) with LLM-friendly hints. Zero deps.

Python port of tool-arg-coerce. Coerce LLM-generated tool args to JSON Schema types. 'true'→True, '5'→5. Zero deps.

LLM prompt templates with Jinja2 syntax. Role-aware Messages builder; output ready for Anthropic / OpenAI SDKs. Python port of agentprompt-rs.

Fill missing tool args from schema defaults before validation. Dict schema or function signature. Nested object + array recursion. Zero deps.

USD cost calculator for Google Gemini API calls. Long-context tiers, thinking tokens, grounding surcharge. Zero deps. Sibling to claude-cost-py + openai-cost-py.

Cache-aware USD cost calculator for OpenAI API calls. Batch API discount, prompt-cache discount, alias resolution. Zero deps. Sibling to claude-cost-py.

Compute the USD cost of a Claude API call from its usage block. Cache-aware, Bedrock-aware (versioned IDs, inference profiles, ARNs), zero dependencies. Python port of claude-cost.

Per-tool call-rate enforcement for LLM agents — sliding window, thread-safe, zero dependencies

Extract JSON, code blocks, lists, and XML tags from LLM text responses — zero dependencies

No description.

No description.

No description.

Per-run budget tracking for tokens, cost, and API calls

No description.

Point-in-time snapshots of agent conversation state for debugging and checkpointing

Fleet-level batch dispatcher for LLM APIs. Pool requests across coroutines, route to provider Batch APIs, save 50% on cost without rewriting your agent loops.

No description.

Parent/child task hierarchy with status tracking and completion rollup

Prompt-cache observability for LLM APIs. Per-call hit ratios, cost saved, regression alerts, miss-aware retry. Anthropic + OpenAI + Bedrock.

Network egress firewall for AI agent tools. Declarative domain allowlist; throws on violation. Optional reqwest-middleware integration.

LLM prompt templates with Jinja2 syntax. Render system/user/assistant turns into a typed message list, ready for the Anthropic or OpenAI SDK.

Hash and version prompt templates so eval results, cache keys, and audit logs stay stable when templates change. Whitespace-normalized SHA-256. Zero deps.

Cost + latency aggregation for LLM agent runs. Group calls into named runs, get totals, p50/p95, and per-model breakdowns. Composes with cachebench.

Snapshot tests for AI agent traces. Record once, replay-and-compare on every run; the agent equivalent of Jest snapshots.

Analyze and simplify JSON Schemas that exceed Anthropic strict-mode compiled-grammar limits. Spots nullable unions, large enums, deep nesting, and applies documented workarounds.

Strip <thinking>/<think> reasoning tags from LLM output. Returns clean text + extracted thinking. Configurable tag set, markdown-style support. Zero deps.

Request coalescing / single-flight for LLM calls. Many concurrent callers same key, one underlying call. asyncio + threading. Zero deps.

VCR-style record + replay for LLM calls. Hash-based matching, RECORD/REPLAY/AUTO modes, sync + async, custom hasher. Zero deps.

Version + content-hash prompt templates over time. Monotonic per-name versions, content-stable hashing, unified-diff between versions, JSON save/load. Zero deps.

Python port of @mukundakatta/agentfit: fit your messages into the LLM context window

Calculate OpenAI API call cost from a usage block. Cache-aware, supports GPT-5, GPT-4.1, o3, o4. Zero deps.

Time-windowed USD/token budget for LLM calls. Enforce per-minute/hour/day caps. Thread-safe. Zero deps. Python port of llm-budget-window (Rust).

Python port of @mukundakatta/agentguard: network egress firewall for AI agents