Quick Jump → 🤖 CodeSlim · 🛡️ Bias Drift Guardian · 🧰 Stack · 📊 Stats · 🤝 Connect
class SachinPaunikar:
role = "Agentic AI Engineer | LLM Infrastructure Specialist | MLOps Engineer"
location = "Nagpur → Open to Remote Global & India Tech Hubs 🇮🇳"
company = "SparrowAI Research and Development Center"
education = "B.E. Electrical Engineer — Nagpur University | CGPA: 8.2 / 10.0"
open_to = [
"Agentic AI Engineer",
"LLM Infrastructure Specialist",
"MLOps Engineer",
"AI/ML Engineer",
"Remote Global · India Tech Hubs ",
]
core_expertise = [
"🤖 Agentic AI Pipelines (LangGraph · Multi-Agent · Tool-Calling · AST Guardrails)",
"⚡ LLM Infrastructure — Token Optimization · Context Engineering · Fallback Chains",
"🔧 Code Quality Automation (LibCST · Radon · Vulture · Tree-Sitter)",
"🏭 MLOps — Docker · FastAPI · Railway · GitHub Actions CI/CD · Model Monitoring",
"📊 Data Drift Detection · Fairness Auditing · Explainable AI (SHAP · DiCE)",
"🗣️ RAG Pipelines · Ollama · LangChain · FAISS · Groq · Generative AI",
]
flagship = "CodeSlim — 76% LLM token reduction · 96/96 tests · $0 cost for 80% of fixes"
currently = "Building production-grade Agentic AI systems & LLM infrastructure tools"
philosophy = "Make AI infrastructure measurably reliable, not just impressively capable."| # | Project | Description | Stack | Demo |
|---|---|---|---|---|
| 🏆 | CodeSlim | NEW · Agentic AI CLI engine — 6-stage LangGraph pipeline. Cuts LLM token bloat by 76% with LibCST, AST guardrails & multi-provider fallback chain | Python 3.11 LangGraph LibCST Radon Vulture FastAPI Ollama Docker |
⭐ Repo |
| 🥇 | Bias Drift Guardian | Real-time AI fairness & drift monitoring. EU AI Act aligned. Intersectional bias detection across compound subgroups | Python Streamlit FastAPI SHAP Fairlearn Docker |
🔴 Live |
| 🥈 | Urban Sound Classifier | 96.63% accuracy on UrbanSound8K. Hybrid U-Net + CNN with real-time mic classification | TensorFlow U-Net Librosa TFLite Flask |
— |
| 🥉 | Transcript → Ad Generator | NLP pipeline: transcript ingestion → NER → LLM ad copy → async video rendering | spaCy Redis Queue MoviePy Docker GitHub Actions |
— |
| 4️⃣ | LLM-PlayBook | RAG pipelines & tool-using agents with retrieval evaluation harness (Recall@k, MRR) | LangChain FAISS Groq FastAPI Docker |
— |
| 5️⃣ | Skin Lesion Segmentation | Medical AI: U-Net pixel segmentation on HAM10000. Temporal tracking with automated >15% growth alerts | TensorFlow U-Net OpenCV Albumentations |
— |
The AI coding explosion created a new problem: structurally bloated, hallucination-prone code.
CodeSlim is the automated guardrail that sits between LLMs and your production codebase.
Industry benchmark reality (2025–2026):
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
• AI-generated code has 1.7× MORE BUGS than human-written code [CodeRabbit 2026]
• 96% of engineers DISTRUST unverified AI code [Sonar / SO 2026]
• 5.2–21.7% of AI package suggestions are NON-EXISTENT [USENIX 2025]
• Average PR size grew 154% LARGER post-AI adoption [Google DORA 2025]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
CodeSlim solution: 76% token reduction · $0 LLM cost for 80% of fixes · 0% hallucination corruption
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
| Capability | What It Does | Impact |
|---|---|---|
| ⚡ 76% Token Reduction | LibCST lossless CST pruning + TF-IDF docstring compression | Massive LLM cost savings |
| 🛡️ AST Safety Gate | ast.parse() + public signature preservation |
0% hallucination corruption |
| 🤖 LangGraph DAG | Stateful multi-node pipeline with conditional branching | Production-grade orchestration |
| 💰 $0 for 80% of Fixes | Deterministic CST node — no LLM needed for import/dead-code purge | Zero API cost baseline |
| 🔒 Local-First Privacy | Ollama (qwen2.5-coder:3b) runs 100% offline on GTX 1650 |
Code never leaves your machine |
| 🎯 3-Tier Classifier | Auto-Safe → Suggest → Flag-Only confidence tiers | Zero unintended breaking changes |
| 💾 SHA-256 Cache | DiskCache stores LLM completions — re-runs cost $0.00 | Complete cost optimization |
| 🚀 GitHub Action | One-line CI integration, zero infrastructure setup | Automated PR audits |
| 🔭 HTML Observatory | Standalone Tokyo Night interactive dashboard + surgery modal | Instant codebase visibility |
# One-liner: Analyze & auto-fix AI bloat in any Python project
codeslim analyze ./src/ --format rich
codeslim optimize ./src/utils.py --apply --backup
codeslim scan ./src/ --export-html observatory_report.htmlBloatScore = min(100.0,
0.30 × CyclomaticComplexity + ← primary driver (Radon CC)
0.25 × NestingDepth + ← AI nesting hell detector
0.20 × DeadCodeLines + ← Vulture-detected bloat
0.15 × CognitiveComplexity + ← Lizard NLOC sensor
0.10 × DuplicationRatio ← MD5 token-hash sliding window
)
| Grade | Score | Meaning |
|---|---|---|
| A | 0–20 | Clean — no LLM intervention needed |
| B | 21–40 | Good — Deterministic Fix Node handles it |
| C | 41–60 | Moderate — LLM Refactor Node invoked on CC > 10 functions |
| D | 61–80 | High bloat — full 6-stage pipeline runs |
| F | 81–100 | Critical — AI over-engineering detected; immediate action required |
Detect bias before it becomes a lawsuit. Monitor drift before it breaks your model.
What makes it unique: Standard fairness tools check one attribute at a time (gender or age). Bias Drift Guardian detects compound discrimination across intersecting subgroups — the kind courts care about.
Standard: "No gender bias detected" ✅ (Male: 70%, Female: 68%)
Ours: "Female employees aged 50+ → only 38% approval rate!" ❌ (Disparity: 0.48)
| Capability | Details |
|---|---|
| 🎯 Intersectional Fairness | Compound subgroup analysis (gender × age × race) |
| 📊 Drift Detection | PSI · KS Test · Chi-Square with configurable thresholds |
| 🔍 Root Cause Analysis | SHAP feature importance drift attribution |
| 🔮 Counterfactual XAI | DiCE What-If explanations (constraint-aware, EEOC-auditable) |
| 📁 Multi-Dataset | German Credit · Adult Census · COMPAS Recidivism |
| 🚀 Deployment | Docker Compose · FastAPI · Streamlit Cloud · MIT licensed |
Core Languages
Agentic AI & LLM Pipelines
Machine Learning & Deep Learning
Responsible AI & Explainability
Data & Visualisation
NLP & Computer Vision
Deployment & MLOps
Two domains. One mission.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🤖 AGENTIC AI ENGINEERING (CodeSlim)
→ AI systems that autonomously audit, minimize, and guard code quality
→ Deterministic-first: 80% of work done by C-native tools, not LLMs
→ LangGraph stateful DAGs · AST guardrails · Multi-provider fallback chains
→ The AI coding explosion demands automated quality enforcement
⚖️ RESPONSIBLE AI (Bias Drift Guardian)
→ What happens AFTER deployment is where most teams go blind
→ 80% of models experience drift within 6 months of going live
→ Compound discrimination (gender × age × race) undetected by
standard single-attribute fairness tools
→ EEOC / EU AI Act compliance is now a legal requirement
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Mission: Build AI that is fast, safe, auditable, and fair.
Areas of deep expertise:
- Agentic AI Pipelines — LangGraph DAG orchestration, multi-node state machines, deterministic tool integration
- AST & CST Engineering — LibCST lossless transformers, Tree-Sitter C-native parsing,
ast.parse()guardrails - Intersectional Fairness — compound multi-attribute subgroup analysis beyond single-attribute tools
- Data Drift Detection — PSI, KS Test, Chi-Square with root-cause attribution via SHAP
- Counterfactual Explanations — DiCE-based What-If analysis, constraint-aware and audit-ready
- Regulatory Compliance — EEOC (US), EU AI Act, GDPR-aware system design
- MLOps — Docker, FastAPI, GitHub Actions CI/CD, Pydantic V2, structured logging
🏢 SparrowAI Research and Development Center | Data Scientist & AI Researcher | Aug 2025 – Jan 2026
→ Built CodeSlim: 6-stage LangGraph agentic pipeline with LibCST + AST guardrails (96/96 tests)
→ Built Bias Drift Guardian: production fairness monitoring system (EEOC / EU AI Act)
→ Intersectional bias detection across compound subgroups (Female + Age 50+ → 38% approval)
→ Drift detection via PSI, KS Test, Chi-Square; SHAP root-cause + DiCE counterfactual XAI
→ Dockerised full stacks; GitHub Actions CI/CD; 96-test green suites on both systems
🏢 Sparrow AI Pvt. Ltd. | Data Science Intern | Jan 2025 – Jun 2025
→ Customer churn prediction & sales forecasting (classification + regression)
→ Automated preprocessing pipelines — reduced manual effort ~30%
→ Stakeholder dashboards (Streamlit · Matplotlib · Seaborn)
🎯 Currently open to:
Agentic AI Engineer · LLM Infrastructure Specialist · MLOps Engineer · AI/ML Engineer
🌍 Available for: Remote Global + India Tech Hubs (Pune · Bangalore · Hyderabad · Mumbai)
If you work in Agentic Systems, LLM Infrastructure, MLOps, FinTech, HealthTech, or AI Tooling — let's talk.
📬 Response time: < 24 hours
"Build AI that is fast, safe, and fair —
with deterministic guardrails at every stage of the pipeline."