---

01 // NEURAL_DOSSIER [IDENTITY_MATRIX]

# OPERATOR IDENTITY SPECIFICATION OPERATOR: Suprovo Mallick CALLSIGN: DYNO ROLE: Generative AI & ML Systems Engineer ACADEMICS: B.Tech in Computer Science @ KIIT (2023 - 2027) ACCREDITATION: BS in Data Science (Foundation) @ IIT Madras COORDINATES: West Bengal, India [88.3639° E, 22.5726° N] # COMBAT MANIFESTO DIRECTIVE: > "Notebook prototypes are child's play. I engineer verifiable, evidence-grounded, low-latency AI pipelines and multi-agent swarms architected to withstand heavy production concurrency." ⚡ TELEMETRY INTEGRITY GAUGES Adaptive RAG & Dynamic Pruning [████████████████] 100% (NLI Filtering) Multi-Agent State Orchestration [██████████████░░] 92% (LangGraph StateGraphs) Deterministic Output Grounding [██████████████░░] 90% (Token-Overlap Bypass) Inference Concurrency & APIs [███████████████░] 95% (FastAPI / 4-bit Quantization) Open Source Runtime Patches [████████████████] 98% (ComfyUI / LlamaIndex / Haystack)

[RUNTIME_STATUS] RETRIEVAL_LATENCY = < 2.0s HALLUCINATION_RATE = < 10% TOKEN_OPTIMIZATION = ~80% PRUNED CITATION_PRECISION = 100% VERIFIED

---

02 // ARCHITECTURAL_CORE [EVIDENCE_FIRST_AI]

"Foundation models are commoditized runtime engines. Real AI engineering begins at query semantics, dynamic pruning, context sanitation, and deterministic output verification."

flowchart TD
    classDef influx fill:#050816,stroke:#00f7ff,stroke-width:2px,color:#00f7ff;
    classDef router fill:#120136,stroke:#7209b7,stroke-width:2px,color:#fff;
    classDef filter fill:#240046,stroke:#ff007f,stroke-width:2px,color:#ff70a6;
    classDef mesh fill:#0a192f,stroke:#00ff99,stroke-width:2px,color:#00ff99;
    classDef gate fill:#1f1147,stroke:#ffd166,stroke-width:2px,color:#ffd166;
    classDef out fill:#003566,stroke:#00e5ff,stroke-width:2px,color:#fff;

    Q[User / Telemetry Stream Influx]:::influx --> R[Semantic Router Node]:::router
    R --> S[Hybrid Dense-Sparse Fetch: ChromaDB + BM25]:::router
    S --> P[NLI Context Pruning Engine: Threshold >= 0.70]:::filter
    P --> M[Multi-Agent Reasoning Graph: LangGraph]:::mesh
    M --> V{Deterministic Gate: Overlap Heuristic >= 0.72?}:::gate
    V -- Bypass Heavy LLM --> O[Async FastAPI REST Endpoint]:::out
    V -- Ambiguous Match --> C[Deterministic Critic Node]:::filter
    C --> O

Loading

LAYER	ENGINE SUBSYSTEM	ARCHITECTURAL IMPACT
01	Hybrid Vector-Sparse Indexing	Combined ChromaDB dense embeddings with BM25 sparse keyword rankers for multi-modal recall.
02	Cross-Encoder NLI Pruning	Strips duplicate and contradictory claims with NLI score $\ge 0.70$, slashing token spend by $\sim 80%$.
03	Stateful Multi-Agent Mesh	LangGraph orchestration isolating Router, Extractor, Synthesizer, and Critic nodes with cyclic recovery.
04	Deterministic Verification Gate	Token-overlap heuristic ($\ge 0.72$) eliminates costly validation LLM calls; bounds hallucinations to $< 10%$.
05	Low-Latency Runtime Delivery	Multi-worker async FastAPI backend with containerized Docker deployment delivering $< 2.0\text{s}$ execution.

---

03 // FLAGSHIP_SYSTEMS [PRODUCTION_NODES]

🧬 NODE 01 // ADAPTIVE EVIDENCE RAG Intelligent, Self-Evaluating Context-Pruning Retrieval Engine Latency: < 2.0s • Token Pruning: ~80% • Hallucination Rate: < 10% • EM Accuracy: > 75% [ARCHITECTURE FLOW] Query ──► Dynamic Expansion ──► 5x Parallel Fetch ──► NLI Pruning (>=0.70) ──► Grounded Qwen-4bit Hallucination Suppression ($&lt; 10%$): Employs strict NLI contradiction filtering and atomic claim extraction prior to synthesis. Context Sanitation: Cross-encoder NLI independence metric eliminates $\sim 80%$ of redundant tokens, ensuring ultra-dense context windows. Full-Stack Container: Containerized React.js + FastAPI pipeline powered by a local 4-bit quantized Qwen model for sub-2-second inference.

🌌 NODE 02 // PLUTO v2 Multi-Agent Document Intelligence & Evidence Verification Mesh Concurrency: 8-Worker Pool • Extraction Latency Drop: ~80% • Citation Precision: 100% [MULTI-AGENT PIPELINE] Document ──► Router Node ──► Extractor + Retriever ──► Synthesizer ──► Critic Gate ──► Verified Payload 100% Verifiable Quote Attribution: Router $\to$ Extractor $\to$ Synthesizer $\to$ Critic state graph guarantees unbacked claims are scrubbed. 80% Latency Drop: Multi-worker thread pools combined with an internal two-tier LLM response cache eliminate redundant computation. Heuristic Bypass Gate: Token-overlap heuristic ($\ge 0.72$) bypasses heavy LLM critic verification for obvious evidence matches.

🧠 NODE 03 // INTENT COMPILER Autonomous Architecture, Schema & Technical Specification Engine Engine: LangGraph StateGraph • Inference: Groq LLaMA-3 LPU • Stack: Streamlit + Python [AUTONOMOUS COMPILATION] Product Vision ──► Agent Graph ──► Requirements Spec + Architecture Blueprints + DB Schemas + Pseudocode Compiles natural-language product ideas into complete, production-grade technical blueprints, database schemas, and executable pseudocode. Employs stateful LangGraph agents that autonomously critique technical viability and validate dependency graphs before compilation.

SYSTEM NODE	CORE STACK & ARCHITECTURE	KEY PERFORMANCE METRIC	LIVE ENDPOINT
🌾 WHEAT GUARDIAN	Edge CV Crop Pathology (EfficientNetV2B2 + OpenCV + FastAPI)	93%+ Accuracy • $\sim 300\text{ms}$ Latency	DEPLOYED APP
🌍 TRANSLATE-V2	Low-Resource NMT Transformer (PyTorch + NLLB-200 + FastAPI)	-38% Latency Gain via Batching	HF SPACE
🥗 AAHAR	Dietary RAG Assistant (LangChain + Gemini + ChromaDB + React)	Personalized Multi-Turn Memory	WEB APP
💰 AI EXPENSE ADVISOR	Financial Retrieval Engine (Streamlit + LangChain + Gemini)	Semantic Spend Classification	GITHUB REPO

---

04 // OPEN_SOURCE_OPS [17+_MERGED_PRS]

PRODUCTION REPOSITORY	IMPACT HIGHLIGHT & PR MISSION	DOMAIN	STATUS
Comfy-Org / ComfyUI (PR #16346 / #16312)	Runtime Reliability: Guarded prompt_worker against unhandled execute exceptions, preventing execution thread crashes.	GenAI Core Engine
open-webui / open-webui (PR #30047 / #30006)	Routing & Observability: Extended parse_custom_headers() to propagate model_id alongside user contexts for downstream routing.	LLM Interface
run-llama / llama_index	Context Integrity: Implemented header-aware deterministic chunking and post-RAG verification primitives.	Agentic RAG
deepset-ai / haystack	Evidence Validation: Created factual consistency and hallucination scoring modules against source documents.	Search Pipelines
MakazhanAlpamys / Soup	Distributed Training: Patched DeepSpeed empty-param-group guards to properly forward transformer model arguments.	Model Training
smoothAPI & Surgite	Backend Resilience: Built async retry predicates, backoff logic, and extensible model provider wrappers.	Cloud / APIs

---

05 // TACTICAL_ARMORY [FULL_SPECTRUM_STACK]

---

06 // INDUSTRY_LOG [EXPERIENCE_&_RESEARCH]

🧪 ROLE // AI CHATBOT & GENAI BACKEND DEVELOPER FED Society | 2024 — PRESENT Client ──► [FastAPI Gateway] ──► [Vector DB RAG] ──► [Memory Matrix] ──► Response Modular FastAPI Microservices: Engineered scalable backends integrating persistent vector databases with multi-turn conversation memory. 30% Latency Reduction: Accelerated end-to-end response pipelines through embedding matrix pooling and selective cross-encoder reranking. Production LLM Endpoints: Architected robust fallback-enabled inference endpoints powering conversational wellness workflows.

📚 RESEARCH // RAAHI (PIC-2026) Real-Time Transport Tracking in Low-Connectivity Cities Proceedings of the 15th Project Innovation Contest (PIC-2026) Satellite Event of ICDCIT-2026, KIIT status: Peer-Reviewed & Published selection: Top 40 out of 160 submissions globally domains: [Distributed Systems, Edge Routing, Telemetry Sync] Formulated a fault-tolerant telemetry synchronization mechanism maintaining positional continuity across packet-loss-heavy cellular networks.

---

07 // ALGORITHMIC_CORE [PROBLEM_SOLVING]

ALGORITHMIC DOMAIN	MASTERY DEPTH	CORE HEURISTICS & PATTERNS
Dynamic Programming & Graphs	[██████████████░░] 88%	State Compression, Topological Sort, Shortest Path (Dijkstra), Memoization
Sliding Window & Two Pointers	[████████████████] 96%	Dynamic Subarrays, Monotonic Queues, Influx Stream Optimization
Trees, Heaps & Tries	[██████████████░░] 90%	LCA via Binary Lifting, Prefix Search Trees, Priority Cache Scheduling
Asymptotic Bounds & Memory	[█████████████░░░] 85%	Space-Complexity Guards, Bit Manipulation, CPU Cache Locality

---

08 // CREDENTIALS_VAULT [CERTIFICATIONS]

ISSUING BODY	SPECIALIZATION / CERTIFICATION	VALIDATION
Udemy	AI Engineer Agentic Track: Complete Agent & MCP Architecture	VERIFY CREDENTIAL
IBM / Coursera	IBM RAG and Agentic AI Professional Specialization	VERIFY CREDENTIAL
AWS / Coursera	AWS Generative AI Applications Specialization	VERIFY CREDENTIAL
AWS / Coursera	AWS Fundamentals Specialization	VERIFY CREDENTIAL
Microsoft / LinkedIn	Career Essentials in Generative AI	VERIFY CREDENTIAL
IIT Madras	Foundation Level Certificate, BS in Data Science & Applications	VERIFY CREDENTIAL

---

09 // TELEMETRY_STREAM [REPO_METRICS]

---

10 // COMM_LINK [TRANSMISSION_PORT]

[SYSTEM_TRANSMISSION]
STATUS   = "LISTENING_ON_ALL_PORTS"
CHANNELS = ["Agentic RAG Architectures", "Multi-Agent System Graphs", "Deterministic AI Verification", "High-Throughput ML Backends"]
QUOTE    = "Logic over guesswork. Evidence over hallucinations. Deployed systems over notebooks."

🏆 GitHub Achievements

Badge	Description	Status
🦈 Pull Shark	Opened pull requests that have been merged	Bronze 🥉 ✅
🎯 YOLO	Merged a pull request without a review	✅ Unlocked
⚡ Quickdraw	Closed an issue or PR within 5 minutes	✅ Unlocked
👥 Pair Extraordinaire	Co-authored commits on merged pull requests	Bronze 🥉 ✅
🧠 Galaxy Brain	Answered discussions with accepted answers	⏳ In progress (8 answers posted)