MIMIC
COGNITIVE USERBOT ENGINE / BEHAVIORAL DIGITAL CLONE
The world's first SaaS platform that designs fully autonomous, conversational digital clones on Telegram. Integrating Amazon Bedrock and semantic RAG on Qdrant, MIMIC acts on your behalf, learning from your chat history to perfectly match your phrasing, scheduling, typos, and personality.
< The Authenticity Paradigm />
Traditional AI Chatbots
Break trust in 5 seconds
- · Instantaneous Replies: Sending complex responses 0.1s after reading, which is physically impossible.
- · Robotic Perfection: Flawless punctuation, overly formal syntax, lacking emojis, conversational gaps, or local slang.
- · 24/7 Availability: Replying at 3 AM with constant, identical enthusiasm, exposing automated agents.
- · Repetitive Phrasing: Reusing rigid, pre-configured LLM templates when prompted with similar topics.
MIMIC Cognitive Engine
Indistinguishable from actual humans
- · Humanizer & Jitter: Custom CPM-based typing calculations, reading delays, and splitting thoughts into separate texts.
- · Personal Style Learning: Automatic extraction of your register (lowercase/normal), emoji preferences, and typical slang.
- · Circadian Integration: Clones follow your time zone, sleeping at night, rising slowly, and delaying replies during work.
- · Semantic Vector RAG: Qdrant matches custom rules using e5 embeddings, seeding LLMs with prompts rather than rigid commands.
< AWS Enterprise Infrastructure />
MIMIC is designed as a highly scalable, secure enterprise SaaS. Each user session is sandboxed in independent ECS containers, securing complete privacy and bypassing Telegram spam detections.
ECS Fargate & Isolated Pods
Every active session runs inside its own isolated micro-container on AWS Fargate. Secrets (StringSessions) are stored encrypted in a PostgreSQL database using AES-GCM-256 with custom keys.
Amazon Bedrock Orchestration
MIMIC communicates through AWS Bedrock APIs to orchestrate Claude 3.5. Native Prompt Caching is utilized, preserving up to 90% of chat contexts and cutting operational token expenses by 80%.
Qdrant Vector RAG Pipeline
Training rules (Custom Rules) are vectorized using e5-small models and saved to AWS Qdrant. Cosine similarities above 0.78 inject contextual snippets as dynamic prompt guidelines.
< Anatomy of Simulation />
Dynamic Timing & Self-Correction
Actual people don't type at a flat speed of 1000 characters per minute. MIMIC Humanizer accurately computes variable entry pauses based on content length.
- Split Message Effects: The engine divides extensive text blocks into 2-3 shorter messages, simulating realistic rapid texting.
- Self-Correction Loops: With a 2% probability, the bot injects a realistic keyboard typo, sending an asterisk correction like
*helloafter a 4s pause. - Dialectical Jitter: Introduces variable delays (4s to 120s) for reading and thinking patterns.
# Calculate humanized typing delay
def calculate_typing_delay(text: str, cpm: int = 180) -> float:
char_count = len(text)
base_time = (char_count / cpm) * 60
jitter = random.uniform(-0.15, 0.25) * base_time
total_delay = base_time + jitter
return max(1.2, total_delay)
# Typo injection model
typo_chance = 0.02
if random.random() < typo_chance:
text = inject_keyboard_typo(text)Circadian Rhythms & Busy Modes
Digital clones follow your active time zones and schedules. You maintain complete control over active intervals.
- Night Sleep Gates: During sleep (e.g., 11:30 PM to 7:45 AM), incoming messages are queued in Redis. Clones wake in the morning and respond naturally.
- Activity Curves: Divided into 4 distinct phases (Morning, Office Hours, Lunch, Evening). Clones respond shorter and less frequently during busy hours.
- Read-and-Walk-Away: Randomly triggers "read and ignore" probabilities for low-priority chats, preventing pushy AI behaviors.
Human-In-The-Loop (HITL) Gateways
Absolute safety and moderation control. The system parses chat flows for sensitive contexts, escalating key messages for manual review.
- Risk Classifiers: Financial requests, offline meetings, crypto credentials, or sensitive chat segments trigger immediate pauses.
- Telegram Admin Bot: Escalates warnings straight to your personal account with a detailed card and 4 quick decision options.
- Triple Fallback Schedule: If you don't respond within 45m, 3h, or 8h, the system fires polite human delays ("sorry, heading into a meeting, catch you later!").
Alex: Hey, let's meet tomorrow at 7 PM for dinner? We can discuss the investment details.
Risk: offline_meeting_request (score: 0.94)
Semantic Chat Fact Memory
Instead of relying on basic context files, MIMIC stores a dynamic facts database for every conversation partner, ensuring factual consistency.
- Fact-Store with Decay: Stores custom metrics (e.g. pets, names, weekly plans) with a TTL decay. Inactive facts slowly fade away.
- Relationship Analyzer: Assesses overall tones to define relationship profiles (Friendly, Intimate, Formal, Hostile) and guides prompt injection.
- Out-Of-Character (OOC) Defense: Regex filters inspect LLM outputs, dropping any AI phrasing like "As an AI model" or "I cannot answer".
FACT DATABASE FOR [ALEX]:
< Interactive Cognitive Sandbox />
Interact with MIMIC in real-time. Text Lilya (our demo clone) directly or choose from the suggested prompt chips. The sidebar panel displays **live cognitive telemetry logs** generated by the active engine.
< Startup Product Roadmap />
< AWS Startup Alignment />
Horizontal Scalability (ECS)
Every active digital clone is highly connection-bound to Telegram API hooks. MIMIC's stateless, Docker-based userbot architecture allows us to dynamically scale up to 10,000+ simultaneous active sessions on AWS ECS Fargate clusters with highly predictable RAM footprints.
Enterprise Bedrock APIs
Accessing Claude 3.5 Sonnet and Haiku via Amazon Bedrock ensures production-level reliability, zero data retention security policies, and optimized latency by leveraging AWS private networks.
AWS Infrastructure Growth
MIMIC's B2B SaaS model presents a highly forecastable and high-margin B2B AWS resource consumer. As user volumes grow, our consumption of Bedrock, Qdrant on ECS, RDS, and ElastiCache scales linearly, securing perfect utility of AWS startup grant programs.