sachin paunikar ImdataScientistSachin

Quick Jump → 🛡️ DataShuddi · ⚡ CodeSlim · 🤖 RetinaFace Pro · ⚖️ Bias Drift Guardian · 🧰 Stack · 📊 Stats · 🤝 Connect

👨‍💻 About Me

class SachinPaunikar:
    role        = "Applied Generative AI & Agentic Systems Engineer | Senior Data Scientist"
    location    = "Nagpur → Open to Remote Global & India Tech Hubs (Bangalore · Pune · Hyderabad · Mumbai) 🇮🇳"
    experience  = "SparrowAI Research & Development Center | Sparrow AI Pvt. Ltd."
    education   = "B.E. in Electrical Engineering — Nagpur University (RTMNU) | CGPA: 8.2 / 10.0"
    
    open_to     = [
        "Applied Generative AI Engineer",
        "Agentic Systems Engineer",
        "LLM Infrastructure & RAG Architect",
        "Senior Data Scientist / MLOps Engineer",
        "Remote Global · Relocation Ready"
    ]

    core_pillars = [
        "🛡️  Sovereign RAG Firewalls (LangGraph · Microsoft Presidio · Qdrant Vector Storage · Zero Cloud Leakage)",
        "⚡  Deterministic Agentic Tooling (LibCST · Tree-Sitter · AST Guardrails · Context Minimization)",
        "👁️  Computer Vision Forensics (RetinaFace · ArcFace · 5-Point Landmark Transformation · Deepfake Evidence)",
        "📊  Production Model Governance (Statistical Drift · PSI · KS-Test · Fairlearn · SHAP Intersectional Auditing)",
        "🏭  Full-Stack AI Deployability (Docker Compose · FastAPI · Streamlit · PostgreSQL PostgresSaver · Pytest)",
    ]

    flagships   = {
        "DataShuddi": "8-Node LangGraph RAG Firewall · 100% PII Recall · 0.00% Residual Leak Rate · 66/66 Pytests",
        "CodeSlim":   "76% LLM Token Reduction · LibCST AST Guardrails · 0% Hallucination Corruption",
        "RetinaFace": "Real-Time Face Detection & Deepfake Verification · ~45ms Latency · HuggingFace Deployed",
        "BiasDrift":  "Intersectional Fairness & Drift Monitoring · PSI / KS Tests · EEOC / EU AI Act Aligned"
    }

    philosophy  = "Make AI infrastructure measurably reliable, sovereignly private, and deterministically auditable."

🚀 Featured Flagship Systems

#	System	Architectural Role	Core Stack	Live & Artifacts
🏆	DataShuddi (3.2.0)	Enterprise RAG Security Firewall · 8-node LangGraph state machine, dual-engine Presidio + Ollama PII redaction, self-correcting reflection loop, Qdrant vector storage	`Python 3.11` `LangGraph` `Presidio` `Qdrant` `FastAPI` `PostgreSQL` `Docker`	⭐ Repo
⚡	CodeSlim	Deterministic Agentic Engine · 6-stage AST optimization engine slashing AI code bloat by 76% via LibCST + Tree-Sitter with 0% hallucination	`Python` `LibCST` `Tree-Sitter` `LangGraph` `FastAPI` `Ollama` `Docker`	⭐ Repo
🤖	RetinaFace Pro	Deepfake Evidence & Biometrics · Real-time dense face localization, 5-point landmark alignment & ArcFace hyperspace verification (~45ms edge latency)	`RetinaFace` `ArcFace` `OpenCV` `TensorFlow` `MLflow` `Docker`	🔴 Live App
⚖️	Bias Drift Guardian	Production ML Drift & Fairness · Intersectional bias detection (compound subgroups), PSI, Kolmogorov-Smirnov, SHAP & DiCE counterfactual XAI	`Python` `Streamlit` `Scikit-learn` `Fairlearn` `SHAP` `Docker`	🔴 Live App

🛡️ Flagship #1 — DataShuddi: Enterprise RAG Security Firewall

Garbage In = Garbage Out. DataShuddi is the sovereign pre-ingestion firewall that neutralizes prompt injections, strips boilerplate, and redacts PII with zero offset drift before embeddings touch your vector database.

🏗️ The 8-Node LangGraph State Machine

  [Raw Untrusted Document (.pdf / .docx / .txt)]
  │
  ├─ Node 1: EXTRACTOR              PyMuPDF / python-docx / PaddleOCR raw parser
  ├─ Node 0: SECURITY SANITIZER     Adversarial prompt injection & jailbreak neutralization
  ├─ Node 2: H/F DETECTOR           Cross-page frequency density boilerplate & disclaimer stripper
  ├─ Node 3: PII DETECTOR           Microsoft Presidio (SpaCy en_core_web_lg) + Ollama Qwen 2.5 Dual Engine
  ├─ Node 4: PII MASKER             Interval-resolved right-to-left substitution (Zero offset drift)
  ├─ Node 5: SCHEMA VALIDATOR       Pydantic v2 validation & compliance quality scoring
  │    │
  │    ├─ [Validation Passed] ───> Node 7: OUTPUT NODE (Sentence Chunking -> mxbai 1024d -> Qdrant Upsert)
  │    │
  │    ├─ [Validation Failed] ───> Node 6: REFLECTOR (Qwen 14b Diagnostic Critique Loop, conf >= 0.70)
  │    │                                    └── (Retry Loop < 3) ───> Rerun Node 1 with tuned delta
  │    │
  │    └─ [Retries Exhausted >= 3] ──> 🚨 HUMAN REVIEW QUEUE (data/human_review/queue.jsonl)
  │
  └─ [Sanitized Payload & Vector Ingested] ──> GDPR Article 17 Purge Ready

⚡ Empirical Benchmark Scorecard

Metric	Score	Production Gate	Status
PII Detection Recall	100.0%	$\ge 95.0%$	✅ PASSED
Composite F1-Score	90.9%	$\ge 90.0%$	✅ PASSED
Clean-Path Latency	< 150 ms / doc	$< 350 ext{ ms}$	✅ PASSED
Residual Leak Rate	0.00%	$0.00%$ (Hard Gate)	✅ PASSED
Automated Test Coverage	66 / 66 Tests	$100%$ Green	✅ PASSED

⚡ Flagship #2 — CodeSlim: Agentic AI Code Quality Engine

The AI coding explosion created structurally bloated, hallucination-prone code. CodeSlim is the automated deterministic guardrail that sits between LLMs and your production codebase.

🏗️ The 6-Stage AST Optimization Pipeline

  INPUT → .py file / directory / GitHub PR Webhook
  │
  ├─ Stage 1 · STATIC SENSORS     Radon (CC) · Vulture (dead code) · Lizard (NLOC) · Tree-Sitter AST · MD5 Duplication
  ├─ Stage 2 · CONTEXT MINIMIZER  LibCST Lossless Pruning · TF-IDF Docstring Compression · tiktoken Budget Enforcer
  ├─ Stage 3 · DETERMINISTIC FIX  LibCST unused-import purge · SHA-256 DiskCache · $0.00 LLM cost (80% of all fixes)
  ├─ Stage 4 · LLM REFACTOR       CC > 10 functions only · Ollama → OpenAI → Groq → CST Fallback (privacy-first)
  ├─ Stage 5 · AST SAFETY GATE    ast.parse() syntax check · Public API preservation · Broken LLM → CST fallback
  └─ Stage 6 · OBSERVATORY UI     Rich Terminal Dashboard · HTML Observatory · FastAPI Web Studio · GitHub PR Bot

Bloat Slashed: Slashes LLM token bloat by 76% using deterministic CST pruning with 0% hallucination risk.
$0 for 80% of Fixes: Pure C-native and AST transformations handle dead imports and scaffolding without invoking expensive LLM tokens.

🤖 Flagship #3 — RetinaFace Pro: Deepfake Evidence & Biometrics

Industry-grade computer vision for real-time dense face detection, 5-point landmark transformation, and ArcFace hyperspace identity verification.

Architecture Trade-off: Feature Pyramid Network (FPN) single-pass localization over slow two-stage RPNs; MobileNet-0.25 backbone delivers ~45ms latency (22 FPS) on edge GPUs.
Spatial 5-Point Alignment: Normalizes eye, nose, and mouth corner coordinates, preventing the ~18% embedding degradation common in unaligned verification pipelines.
ArcFace Hyperspace Separability: Employs Additive Angular Margin Loss for compact intra-class clustering and maximized inter-class margins for deepfake and forensic verification.
MLflow Observability: Live tracking of inference latency, confidence calibration, and landmark drift.

⚖️ Flagship #4 — Bias Drift Guardian

Detect bias before it becomes a lawsuit. Monitor drift before it breaks your model.

Intersectional Fairness: Detects compound subgroup discrimination (e.g. Female + Age 50+ $ o$ 38% approval rate) where single-attribute tools report false passes.
Statistical Drift Engines: Automated Population Stability Index (PSI), Kolmogorov-Smirnov, and Chi-Square distribution tests.
Explainability: SHAP feature drift attribution + DiCE counterfactual What-If audit scenarios (EEOC & EU AI Act aligned).

🧰 Tech Stack

Agentic AI, LLM & RAG Systems

Machine Learning & Computer Vision

Data Infrastructure & Databases

Deployment, MLOps & APIs

📊 GitHub Engineering Telemetry & Activity Graph

📈 Experience Highlights

🏢 SparrowAI Research and Development Center  |  Data Scientist & AI Researcher  |  Aug 2025 – Jan 2026
   → Built DataShuddi: 8-node LangGraph RAG security firewall with Presidio + Ollama (66/66 Pytests)
   → Built CodeSlim: 6-stage LangGraph AST engine slashing AI code bloat by 76% (96/96 Pytests)
   → Built Bias Drift Guardian: Intersectional fairness & statistical drift detection pipeline
   → Deployed containerized Docker Compose stacks and automated GitHub Actions CI/CD workflows

🏢 Sparrow AI Pvt. Ltd.  |  Data Science Intern  |  Jan 2025 – Jun 2025
   → Customer churn prediction & sales trend forecasting (classification + regression)
   → Automated preprocessing and ETL pipelines — reduced manual effort by 30%
   → Engineered interactive analytical telemetry dashboards (Streamlit · Power BI)

🤝 Let's Connect

🎯 Currently open to:
Applied Generative AI Engineer · Agentic Systems Engineer · LLM Infrastructure Specialist · Senior Data Scientist

🌍 Available for: Remote Global + India Tech Hubs (Bangalore · Pune · Hyderabad · Mumbai)

📬 Response time: < 24 hours

"Build AI infrastructure that is fast, sovereignly private, and deterministically auditable at every stage of the pipeline."

Navigation Menu