Author & Engineer: Harsh Ramesh Nerkar
Repository:harshrameshnerkar/Agentic-AI
Program: Agentic AI β Production Engineering Curriculum
This repository contains the complete 20-day hands-on engineering assignments completed across the 20-Day Autonomous Agentic AI Engineering Curriculum.
Moving far beyond naive prompt wrappers and single-turn completions, this repository implements the full architectural spectrum of modern Autonomous Agent Engineering, Multi-Agent Swarms, Model Context Protocol (MCP), and LLMOps at Scale:
- Deterministic ReAct execution loops & structured tool schemas
- Multi-agent hierarchical supervisor patterns & worker swarms
- Adversarial prompt injection defense & safety guardrails
- Sub-millisecond intent routing & hybrid adaptive RAG
- High-throughput asynchronous concurrency (
asyncio.gather) - Durable transactional state machines with SQLite Write-Ahead Logging (WAL)
- Human-in-the-Loop (HITL) approval gates with cryptographic SHA-256 audit chaining
- Automated CI/CD regression gates with golden evaluation suites
- Production containerization (Docker & docker-compose) and real-time observability
| Days | Topic Name | Technical Scope & Key Deliverables | Code Link |
|---|---|---|---|
| 1 | LLM Fundamentals & API Basics | β’ Setup, first LLM API calls, tokenization mechanics β’ Temperature, top_p, frequency penalty parameters, context windows, cost models |
day_01/ |
| 2 | Advanced Prompting & Structured Output | β’ Chain-of-Thought (CoT), Task Decomposition, Few-Shot prompting β’ Pydantic schema validation, JSON mode parsing, prompt templates & versioning |
day_02/ |
| 3 | Embeddings, Vector DB & RAG | β’ Dense vector embeddings, cosine/dot similarity, chunking strategies β’ Vector database indexing, retrieval failure modes, semantic search |
day_03/ |
| 4 | Tool Calling & Function Execution | β’ OpenAI/Anthropic/Gemini function schemas, tool registration β’ Multi-step tool use, argument extraction, error handling loops |
day_04/ |
| 6 | Tool Calling & MCP | β’ Model Context Protocol (MCP) integration & tool design best practices β’ Blast-radius classification, safe read vs destructive write tiers, compensating rollbacks |
day_06/ |
| 7 | Agents: ReAct, Frameworks & Memory | β’ Pure Thought-Action-Observation ReAct execution loops β’ LangGraph StateGraph workflows, conversational & episodic checkpointed memory |
day_07/ |
| 8 | Multi-Agent Systems & Orchestration | β’ Centralized Supervisor router, collaborative worker swarms β’ Inter-agent communication protocols, context compaction & engineering |
day_08/ |
| 9 | Guardrails, Security & Agent Evaluation | β’ Prompt injection detection, jailbreak mitigation, Canary token leaks β’ Dual-LLM safety boundary guards, adversarial red-teaming test suite |
day_09/ |
| 10 | Capstone, Deployment & Final Review | β’ OpsSentinel AI Midterm Capstone: autonomous multi-tool SRE diagnostic copilot β’ Containerized deployment, interactive Streamlit UI, comprehensive review |
day_10/ |
| 11 | Advanced Retrieval Architectures | β’ Hybrid dense/sparse search (BM25 + vector), graph & structured metadata RAG β’ Query rewriting, sub-query decomposition, dynamic retrieval routing |
day_11/ |
| 12 | Fine-Tuning & Model Customisation | β’ Decision framework: RAG vs Fine-Tuning, SFT data synthesis & QC filters β’ Parameter-Efficient Fine-Tuning (PEFT/LoRA), multi-tier cost router |
day_12/ |
| 13 | Production Agent Architecture | β’ High-throughput async tool dispatch via asyncio.gather (3.2x speedup)β’ SQLite WAL ACID checkpointer, multi-tenant state isolation, circuit breakers |
day_13/ |
| 14 | LLMOps, Evaluation at Scale & Incidents | β’ Automated GitHub Actions CI regression gates with 100-case golden datasets β’ Canary/shadow trace evaluators, emergency kill switches & postmortems |
day_14/ |
| 15 | Production Capstone & Final Assessment | β’ Hardened autonomous SRE control center with sub-2ms intent router β’ FastAPI microservice, SHA-256 Merkle audit trail, HITL approval gate β’ 4-Page publication TDD, 10x cost projection, 20-min executive defense |
day_15/ |
| 16 | Scoping a Real Problem | β’ Session 1 (Discovery): Stakeholder problem statement in stakeholder's words, 7-stage workflow replacement analysis & Monte Carlo MTTR simulation β’ Session 2 (Constraints): Signed-off success metrics charter (95% accuracy, <=90s p95 latency, <=$0.15 cost ceiling), 3-tier blast radius classification, 5 Never-Automate red lines & cryptographic HMAC-SHA256 HITL gateway β’ Session 3 (Feasibility Spike): Data & documentation audit, 2-hour timeboxed retrieval spike (BM25 vs. Dense vs. Hybrid RRF), official Feasibility Note & 93% static retrieval ceiling analysis β’ Session 4 (Plan, Estimate & Eval Set First): Filled MoSCoW Sprint Planning Board (21 Must, 13 Should, 8 Could SPs), team capacity balancing (29h / 32h), pre-code 30-case golden evaluation dataset with 6 adversarial red-line scenarios |
day_16/ |
| 17 | Build Sprint 1 | β’ Session 1 (Standup & Riskiest Task): 15-minute standup (Done, Next, Blocked), Sprint 1 Kanban tracker moving TASK-1.1 to In Progress, concurrent async diagnostic tool dispatcher, 50k-line log anomaly compactor & circuit-breaker resilience β’ Session 2 (Simplest Thing That Works): Non-agent baseline evaluation (Plain Prompt 0% vs. Vanilla RAG 30%), golden 20-case evaluation dataset, empirical proof of ephemeral cluster gap β’ Session 3 (First Real Iteration): Scientific step-by-step evolution (+Retrieval 30%, +Cluster Tool 100%, +Conditional Routing 100% with 13.7% latency reduction) β’ Session 4 (Progress Update): Strict 5-line executive status update to mentor Dr. Rostova & Redis cluster state caching sign-off |
day_17/ |
| 18 | Build Sprint 2 & Code Review | β’ Session 1 (Standup & Replan): 15-minute standup & explicit scope cut (Slack bot moved from Must to Could), capacity rebalancing (13.5h load vs 16.0h cap), updated Sprint Board & Replan Manager β’ Session 2 (Integration & Safety): Universal REST/CLI delivery surface, regex PII scrubber (SSN, CC, email, JWT, API tokens), blast radius classifier, HMAC approval gateway & tamper-evident SHA-256 hash-chained audit logging β’ Session 3 (Code Review): Formal PR #18, line-by-line review addressing 5 comments without defensiveness (UsedNonceRegistry replay defense, ReDoS guards, Enum tiers, RLock thread-safety, @dataclass TriageResult) β’ Session 4 (Regression Suite): Automated one-command CI regression runner evaluating 35 golden cases (30 baseline + 5 regression bugs), 100.0% pass rate, sub-50ms latency, zero-tolerance non-zero exit gate |
day_18/ |
| 19 | Hardening, Documentation & Handover | β’ Session 1 (Standup & Freeze): 15-minute standup & formal feature freeze agreement (v1.0.0-rc1), zero-new-feature invariant, allowed/disallowed commit matrix & automated freeze auditor β’ Session 2 (Break Your Own System): 6-vector adversarial chaos test harness (prompt injections, malformed JSON schemas, empty RAG zero-matches, downstream socket crashes, upstream 429 outages, runaway recursive loops) & 8-column failure-mode matrix β’ Session 3 (Handover Pack): Complete 8-section production engineering manual (1-click setup, architecture, prompt catalog, on-call runbooks, unit economics model with $0.0028/query baseline, roadmap & walkthrough) β’ Session 4 (Clean-Clone Test): Independent peer review sandbox trial by Senior SRE Alex Chen with 0 author questions, resolution of 3 documentation bugs, and AST-verified headless non-interactive execution |
day_19/ |
| 20 | Demo, Retrospective & Final Assessment | β’ Session 1 (Stakeholder Demo): 20-minute executive presentation to VP of Infra & Lead SRE, leading with workflow impact ($42k/mo saved, 45ms MTTR), live incident walkthroughs & explicit refusal boundary disclosures β’ Session 2 (Results & Limitations): Formal 3-page technical results report evaluating stratified pass rates (100.0% on 35 cases), p95 latency (48.2ms), unit cost ($0.0028), scope cut defense & honest limitations β’ Session 3 (Retrospective): Engineering sprint retrospective analyzing estimation variance (+60% on retrieval spike, -100% on cut Slack bot) & deep-dives into 3 mistakes (HMAC replay nonces, ReDoS, and audit lock race condition) β’ Session 4 (Final Assessment & Conversion): 4-Week technical interview rubric (5.0/5.0 Exceeds Expectations), live multi-region Byzantine audit system-design defense, and unanimous Full-Time Conversion Approval |
day_20/ |
The program culminates in OpsSentinel AI Enterprise (day_15/) β an autonomous Site Reliability Engineering (SRE) copilot designed to detect, triage, and remediate distributed cloud infrastructure incidents under enterprise SLAs and strict blast-radius controls.
+-------------------------------------------+
| Inbound Operations Query |
| (Terminal CLI / REST API / Webhook) |
+---------------------+---------------------+
|
v
+-------------------------------------------+
| Sub-2ms Semantic Query Router |
| (Regex Pre-Filter + Tenant-Scoped SOPs) |
+---------------------+---------------------+
|
+--------------------+--------------+--------------+---------------------+
| | | |
v v v v
[ INFO_SOP (RAG) ] [ DIAGNOSTIC_READ ] [ DESTRUCTIVE_WRITE ] [ ADVERSARIAL_ATTACK ]
Vector Runbook SOP Parallel async tools Tier 3 Blast Radius Immediate Abort
Direct response (Metrics, Logs, Topology) State: AWAITING_APV Security Audit Log
| | | |
+--------------------+-----------------------------+---------------------+
|
v
+---------------------------------------------------------------+
| Transactional Durable Checkpointer |
| (SQLite WAL ACID - Crash-Safe State) |
+-------------------------------+-------------------------------+
|
+--------------------+--------------------+
| |
v v
+-------------------------------+ +-------------------------------+
| Human-In-The-Loop Gateway | | Cryptographic Audit Trail Log |
| (HMAC Token / Role Approval) | | (SHA-256 Tamper-Evident Chain)|
+---------------+---------------+ +-------------------------------+
|
v (Approved)
+-------------------------------+
| Executing Blast-Radius Action |
| (Rollback Hook + Verification)|
+-------------------------------+
-
Parallel Concurrency: 3.2x latency reduction over sequential tool dispatch using
asyncio.gatherfor cluster telemetry, logs, and topology checks. - Crash-Resilient State: 100% state recovery across mid-execution crashes with SQLite WAL checkpointing and idempotent replay.
-
Cryptographic Blast-Radius Control: Any Tier-3 destructive operation (service restarts, scaling, cache flushes) is paused in
AWAITING_APPROVAL. Authorized actions produce an immutable SHA-256 Merkle-style audit log. -
CI/CD Quality Gates: Fully wired into GitHub Actions (
.github/workflows/day15_capstone_ci.yml) enforcing$\ge 90%$ pass rate, 0 critical regressions, and sub-250ms p95 latency.
Empirical test data measured on Python 3.12 running against the golden evaluation harness (12 stratified incident cases):
| Metric | Target SLA | Measured Baseline | Status |
|---|---|---|---|
| Overall Pass Rate | PASS | ||
| Critical Regressions | PASS | ||
| Adversarial Jailbreak Defense |
|
PASS | |
| Sequential Diagnostics Latency | β | Baseline | |
| Async Parallel Latency | PASS | ||
| p50 Latency (Overall) | PASS | ||
| p95 Latency (Overall) | PASS | ||
| Audit Chain Cryptographic Integrity | 100% Intact | SHA-256 Validated | PASS |
git clone https://github.com/harshrameshnerkar/Agentic-AI.git
cd Agentic-AI
# Create virtual environment
python -m venv .venv
# On Windows:
.venv\Scripts\activate
# On Linux/macOS:
source .venv/bin/activate
# Install dependencies
pip install -r day_15/requirements.txtpython day_15/test_day_15_all.pyExpected: 16 integration tests passing in ~4 seconds with 100% evaluation pass rate.
python day_15/session_1/api_server.pyMicroservice starts at http://localhost:8000. Access Swagger UI docs at http://localhost:8000/docs.
# Verify cryptographic audit chain integrity
python day_15/session_1/main.py --audit-check
# Run an infrastructure diagnostic query
python day_15/session_1/main.py -q "check cpu and error rate on auth-service"
# Trigger a destructive write requiring HITL approval
python day_15/session_1/main.py -q "restart pod auth-service-prod-0"Harsh Ramesh Nerkar
Autonomous AI Systems Engineer | Production LLMOps Architect
- GitHub: @harshrameshnerkar
- Repository: harshrameshnerkar/Agentic-AI