Autonomous intelligence.
Rooted in character,
bounded by safety.
Project Ishita Goyal (@ishita._g_) introduces an architectural breakthrough in autonomous AI personas. By replacing prompt wrappers with authoritative relational databases, dynamic social graphs, official Meta Graph API protocols, and zero-dollar multi-model routing across Gemma 4 and Gemini 3.5, Ishita achieves authentic human-level connection without hallucination or safety compromise.
Latest releases
Newsroom
Read the latest essays, architectural releases, sociolinguistic research, and engineering breakthroughs behind Project Ishita Goyal.
Introducing Character-State Architecture: Decoupling LLM Cognition from Relational Identity
Why prompt wrapping inevitably leads to character degradation over extended conversation horizons, and how our database-anchored social state guarantees non-hallucinatory identity permanence.
How We Achieved $0.00 Cloud Spend Across 14,400 Daily Conversations
Engineering a zero-cost intelligent routing matrix balancing Gemma 4 26B and Gemini 3.5 Flash Lite under strict rate limits.
The 5-Tier Gated Autonomy Model: Preventing Rogue Behaviors
Deterministic protection against jailbreaks, boundary probing, and unconstrained autonomous social publishing.
Jaipur-Delhi Sociolinguistic Grounding: Engineering Natural Hinglish
Eliminating robotic AI markers and modeling authentic bilingual code-switching for young urban creators.
Why simplistic prompt wrappers fail — and how Character-State Architecture succeeds.
Traditional AI roleplay systems cram extensive character backstory into the LLM system prompt. Over long conversations, context drift erodes facts, hallucinations corrupt boundaries, and agents easily fall victim to prompt injection or repetitive robotic phrasing.
Authoritative Relational Memory
Facts, relationship depth, boundaries, and episodic memories do not live in volatile prompt tokens. They reside in structured SQL databases and 3,072-dimensional vector stores with explicit CANDIDATE, CONFIRMED, and DEPRECATED states.
- Resolves memory contradictions deterministically (e.g. coffee preferences)
- Eliminates memory leakage across different social circles
- Recency, importance, and closeness weighted retrieval
11-Tier Social Graph & Dynamic Boundaries
Every individual in Ishita's orbit is categorized across 11 relationship tiers (from Stranger and Follower to Close Friend and Ex). Each tier strictly enforces disclosure ceilings, conversational warmth, and initiation preferences.
- Exes are strictly bounded: brief, polite, zero romantic ambiguity
- Strangers promote dynamically to Friends after sustained positive contact
- Protects private personal stories from casual followers
Gated Autonomy & The 5-Level Security Shield
Autonomy is not binary. Ishita's independent decision-making operates through a 5-level security hierarchy combined with an immutable, software-enforced global kill switch and anti-double-texting rules.
- Zero double-texting: Will never follow up if contact hasn't replied
- 3-hour per-person cooling interval prevents notification fatigue
- One-click hardware-level kill switch immediately blocks external outputs
Zero-Cost Intelligent Multi-Model Routing
Operates entirely on Google AI Studio's verified free-tier models with sliding-window quota management. High-volume conversational tasks route to Gemma 4 26B (14.4K daily limit), while 250K TPM video tasks leverage Gemini 3.5 Flash Lite.
- Guaranteed $0.00 cloud expenses with strict CostGuard
- Real-time RPM and RPD quota counters prevent 429 errors
- Multi-candidate automatic fallback on service hiccups
Inside the 14-Step Execution Pipeline
Every inbound DM or comment passes through 14 deterministic stages before a single token is transmitted back to the user. Click any stage below to inspect its internal logic and telemetry.
1. Normalization & Slang Extraction
Inbound messages are scrubbed for excessive whitespace, normalized, and scanned for everyday urban Hinglish slang (e.g., 'kya scene', 'sahi mein', 'tapri', 'arre').
Free-Tier Model Allocation & Google AI Studio Rate Limits
Our intelligent ModelRouter maps conversational and perception tasks to the most suitable free-tier model, continuously balancing daily limits (RPD) and context capacity (TPM).
| Model Identifier | Assigned Role | Daily Limit (RPD) | Rate Limit (RPM) | Context Window (TPM) | Cost / Usage Strategy |
|---|---|---|---|---|---|
|
Gemma 4 26B
Primary Chat
|
High-frequency DMs, Hinglish banter, persona dialogue | 14,400 / day | 30 RPM | 16,000 TPM | $0.00 Free Tier • Workhorse |
|
Gemma 4 31B
Reasoning
|
Complex social graph reflection & chat fallback | 14,400 / day | 30 RPM | 16,000 TPM | $0.00 Free Tier • Secondary |
|
Gemini 3.5 Flash Lite
Vision & Video
|
Reel video QA, Instagram photo OCR, Pydantic memory extraction | 500 / day | 15 RPM | 250,000 TPM | $0.00 Free Tier • High-Context |
|
Gemini 3.1 Flash Lite
|
Multimodal reserve & structured parsing backup | 500 / day | 15 RPM | 250,000 TPM | $0.00 Free Tier • Hot Reserve |
|
Gemini 3.8 Flash
Flagship
|
Deep visual aesthetic scoring & high-reasoning tasks | 20 / day | 5 RPM | 250,000 TPM | $0.00 Free Tier • Conserved Quota |
|
Gemini Embedding 2
Vector Memory
|
3,072-dimensional embeddings for episodic vector search | 1,000 / day | 100 RPM | 30,000 TPM | $0.00 Free Tier • Vector Store |
Advancing Autonomous Alignment
Our research focuses on foundational character safety, deterministic relational memory, real-time multimodal perception, and mathematical frameworks for long-horizon identity stability.
Research Working Groups
All publications include formal methodology derivations, mathematical equations, reproducible test benchmarks, and downloadable BibTeX citations.
Live Benchmarks & Experimental Validations
Data gathered from automated test suites, real-time stress testing, and longitudinal conversational evaluations.
Anti-AI Cliché Sanitization
Standard LLMs frequently inject patronizing phrasing ("As an AI...", "Certainly!"). Our ResponseValidator completely purges robotic phrasing while maintaining natural Hinglish grammar.
Continuous Quota Resilience
By dynamically routing between Gemma 4 26B (14.4K RPD) and Gemini 3.5 Flash Lite (500 RPD), the system experienced zero 429 quota rejections and maintained $0.00 billing.
Boundary Integrity
In 500 adversarial dialogue trials attempting to elicit intimate romantic disclosure or personal location from restricted tiers, the system executed zero boundary breaches.
Architectural Tradeoffs: Transparent Pros & Cons
In the spirit of honest scientific evaluation, we document both the breakthroughs and the inherent constraints of our design choices.
Key Advantages (Pros)
-
Strict $0.00 Cloud Expenditure Operates completely within Google AI Studio's free tier with zero risk of unexpected billing surges.
-
Authoritative Identity & Contradiction Resolution Memories do not vanish when context windows compress; historical preferences update cleanly without hallucination.
-
Meta Graph API Compliance Built solely on official OAuth 2.0 and Graph API protocols with zero account-ban risk from browser scrapers.
-
Non-Bypassable Safety Net Hardware kill switch and anti-double-texting guarantees the persona will never spam or harass contacts.
Constraints & Tradeoffs (Cons)
-
!Flagship Model Quota Ceilings Gemini 3.8 Flash is capped at 20 RPD on the free key, requiring automated fallback to Gemini 3.5 Flash Lite for high-volume video analysis.
-
!14-Step Pipeline Latency Overhead Multi-stage classification, memory reranking, and sanitization add ~300ms overhead compared to naive unguided LLM streaming.
-
!Relational Database Dependency Requires an active SQLite or PostgreSQL database instance to persist graph metrics, making purely stateless deployment impossible.
-
!Strict Boundaries May Restrict Open-Ended Play Users seeking unconstrained fictional roleplay will encounter deliberate boundaries with non-intimate relationship tiers.
Experience Ishita's Conversational & Visual Intelligence
Test real-time conversation or inspect how Gemini 3.5 Flash Lite analyzes real Instagram posts from Ishita's feed.
Real-Time Pipeline State
Every interaction through this sandbox executes the live 14-step pipeline on our local server.
- Model: gemma-4-26b-a4b-it • 14.4K RPD
- Billing: $0.0000 USD (Guaranteed Free)
- Anti-AI Cliche Filter: Armed
- Timezone: Asia/Kolkata (IST Clock Active)
- Intent & Emotion Parser: Operational
Click any post below to inspect its AI visual breakdown, aesthetic score, and extracted color harmony.
1st Post — Jaipur Hillside Sunset
Aesthetic Fit: 0.94 • Golden hour cafe view
2nd Post — Coffee & Morning Journal
Aesthetic Fit: 0.96 • Specialty pour-over ritual
3rd Post — Visual Design & Typography
Aesthetic Fit: 0.92 • Design student layout study
4th Post — Pink City Historic Arches
Aesthetic Fit: 0.95 • Architectural curiosity
5th Post — Candid Studio Snapshot
Aesthetic Fit: 0.91 • Relatable studio work
CLI Tooling & API Endpoints
Project Ishita Goyal includes robust developer diagnostics, CLI chat modes, and complete REST endpoints.