Shrikant Lambe · Data-Analytics Engineering Leader

11 / hr-copilot

AI Agent · Recruiting Analytics

HR Analytics Copilot — a dashboard that writes its own SQL.

A recruiting-funnel dashboard with a dual-path AI assistant: numeric questions are answered by an LLM writing SQL that runs client-side against DuckDB-WASM, returning both the number and the query that produced it; interpretive questions are answered from a findings knowledge base with citations. There's no backend database — data lives in local CSVs — and the LLM key stays inside a single server-side API route, never reaching the client bundle. Ships with a full demo mode so the live app works end-to-end with zero key and zero flaky-network risk.

🔍 SQL-writing assistant + citation-grounded insights · zero backend DB · demo mode needs no API key

2

Answer paths — SQL + grounded

0

Backend database

Live

Demo mode, no key needed

Next.jsTypeScriptDuckDB-WASM ClaudeTailwind CSSVercel

Live app ↗ github ⎆

hr-ai-chatbot · recruiting dashboard

ASSISTANT · DATA QUESTION

What's our offer rate for valid Engineering roles?

18.9% — computed live via generated SQL against 37 valid jobs (test/intern postings excluded).

◆ SQL used

SELECT round(100.0 * sum(offer) / count(*), 1) FROM applications WHERE job_valid = true

37

Valid jobs

5

Canonical stages

0

Backend DB

12 / barista

AI Agent · Conversational

Coffee Barista Agent — an ADK agent that won't recommend off-menu.

A conversational ordering assistant built on Google's Agent Development Kit, served through a Streamlit chat UI. The agent is grounded in an actual menu via a get_menu tool call — it asks one clarifying question when a request is vague (hot or cold, sweet or strong) and respects allergen and dietary tags instead of free-associating a drink. A companion script embeds the menu with Vertex AI text-embedding-005 into Firestore, laying the groundwork for semantic menu search.

☕ Menu-grounded recommendations · allergen-aware · Vertex AI embeddings for semantic search

1

Menu-grounded tool call

0

Off-menu recommendations

Vertex AI

Embeddings seed script

Google ADKGeminiStreamlit FirestoreVertex AI Embeddings

Live app ↗ github ⎆

coffee-barista-agent · chat

ORDER · IN PROGRESS

Something cold and not too sweet, and I'm dairy-free.

Iced Americano or the Cold Brew — both dairy-free, neither is sweet. Want oat milk in either?

✓ Grounded in menu.json

Every suggestion checked against real items, prices, and allergen tags — nothing recommended off-menu.

07 / sentinel

Multi-Agent · DataOps

Pipeline Sentinel — a self-healing agent for Airflow.

Five specialized agents — Monitor, Diagnosis, Blast Radius, Remediation, Reflection — coordinate through shared state to diagnose and resolve pipeline failures without paging a human. Pattern memory prevents repeat incidents. A confidence gate auto-remediates only low-risk, high-signal cases; everything else escalates with a full audit trail. Independent build — the numbers below are from its own test/demo scenarios, not a company-monitored production system.

⚡ Demo scenario: resolves failures in ~11s at 94% confidence · escalates everything else with a full audit trail

5

Agents in ReAct loop

94%

Confidence (demo run)

11s

Time-to-resolution (demo)

Claude SonnetLangGraphLangSmith FastAPIAirflowStreamlit

github ⎆ substack ⎆ medium ⎆ linkedin ⎆

pipeline-sentinel · agent dashboard

DEMO SCENARIO · customer_churn_pipeline · transform_features

✓Monitor · task failure detectedT+04s

✓Diagnosis · upstream schema drift on `plan_tier`T+07s

✓Blast Radius · 3 downstream tasks identifiedT+08s

●Remediation · applying `reload_schema` strategyT+09s

✓ Self-healed · confidence 94%

Schema reloaded, 3 tasks re-queued. No human intervention needed in this demo scenario.

08 / loanlens

FinTech · Data Platform

LoanLens — portfolio intelligence for a simulated Series C lender.

10,000 loans, 1.35M payment events, 3 SPVs. Modeled in dbt, reconciled end-to-end to <0.1% tolerance, narrated by Claude. SPV covenant monitoring, 36-cohort vintage analysis, anomaly detection, and investor-grade PDF memo export. Designed and shipped as a two-day sprint.

📊 1.35M payment events · 3 SPVs reconciled to <0.1% · investor-grade PDF memo in one click

1.35M

Payment events

<0.1%

Recon tolerance

36

Cohort vintages

Claude SonnetDuckDBdbt Core SnowflakeStreamlitPython

Live app ↗ github ⎆

loanlens · spv-b portfolio snapshot

SPV-B · Q4 2025 · AS-OF 2026-04-22

$284M

AUM

+3.1% QoQ

3.2%

Default rate

+40bps

0.04%

Recon delta

within tol.

◆ Investor memo · claude

SPV-B at 87% facility utilization — recommend 60-day drawdown pause. 2023 vintage cohorts outperforming 12-month default curve by 40bps.

06 / growth-agent

Agentic AI · SaaS Analytics

Growth Intelligence Agent — a virtual revenue analyst.

Autonomously monitors SaaS growth metrics across 6 categories (new logo, expansion, retention), detects anomalies with severity scoring, and surfaces RAG-grounded strategic recommendations from company playbooks — all through a natural-language interface. Thinks like an analyst, cites like a playbook.

🤖 Surfaces playbook-grounded recommendations autonomously across 6 metric categories

6

Metric categories

FAISS

Playbook retrieval

NL

Query interface

Claude 3.5 SonnetLangChainFAISS RAGStreamlitPlotly

Live app ↗ LinkedIn article → Medium ↗ Substack ↗ github ⎆

growth-intelligence-agent · week 17

SAAS PORTFOLIO · WEEK 17 · REAL-TIME

$2.4M

MRR

+6.2% WoW

118%

NRR

above target

↑2.1%

Churn risk

anomaly

⚠ ANOMALY · HIGH — enterprise churn spiked +2.1pp WoW · 3 accounts flagged

Why did enterprise churn jump this week?

Per ICP playbook §3.2: accounts with <60% feature adoption + no QBR in 90d = highest signal. 3 flagged accounts match both. Recommend CSM outreach this week.

05 / cortex

Snowflake · Enterprise AI

AI-Native Data Platform on Snowflake Cortex.

A reference architecture and live demo for running governed AI natively inside the Snowflake Data Cloud — Cortex LLM functions (COMPLETE, SENTIMENT, CLASSIFY), Cortex Search, and Cortex Analyst, with live interactive demos. Zero data egress is a design property of the in-warehouse pattern, not a third-party-audited guarantee — this is a personal architecture blueprint, not a deployed enterprise system.

🏗 Architecture blueprint: zero data egress by design · end-to-end governed AI patterns · live interactive demo on GitHub Pages

3

Cortex layers

0

Data egress (by design)

E2E

Governed AI pattern

Snowflake CortexCortex Search Cortex AnalystAnthropic APISQLGitHub Pages

Live architecture ↗ Medium → Substack ↗ github ⎆

snowflake-cortex-architecture

GOVERNED AI · IN-WAREHOUSE EXECUTION

-- customer feedback analysis inside Snowflake
SELECT
  customer_id,
  SNOWFLAKE.CORTEX.SENTIMENT(support_notes) AS sentiment,
  SNOWFLAKE.CORTEX.CLASSIFY(
    support_notes, ['billing', 'outage', 'feature']
  ) AS category,
  SNOWFLAKE.CORTEX.COMPLETE(
    'llama3-70b',
    CONCAT('Summarize: ', support_notes)
  ) AS summary
FROM customer_feedback
WHERE region = 'EMEA';

3

Cortex layers

0

Egress

E2E

Governed

03 / retail-pipeline

Data Engineering · Streaming

Real-Time Retail Sales Pipeline — Kafka to warehouse, E2E.

Production streaming platform: Kafka producer → PySpark Structured Streaming → Snowflake → dbt incremental MERGE → Airflow orchestration with automated data quality gates. Fully containerized via Docker Compose. System design documented as an architecture reference.

🔄 Processes 25 events/sec end-to-end · zero DQ failures · fully containerized with Docker Compose

25/s

Kafka throughput

0

DQ failures

E2E

Stream to warehouse

Apache KafkaPySparkSnowflake dbtAirflowDocker

System design → Medium ↗ Substack ↗ github ⎆

retail-pipeline · airflow dag

PIPELINE STATUS · retail_streaming_pipeline

✓Kafka Producer · 25 events/seclive

✓PySpark Streaming · dedup + windowdone

✓Snowflake write · raw + agg layersdone

●dbt run · fact_sales incremental MERGErunning

24.8K

Rows today

$94K

Revenue

0

DQ fails

02 / mktg-ai

GenAI · Enterprise

Marketing AI Intelligence Engine — board-ready in 30 seconds.

Inspired by real AI work at Workday. 4-layer GenAI architecture — Semantic → Context/RAG → Insight Engine → LLM Generation — that transforms raw campaign data into board-ready executive summaries. No analyst in the loop required.

⚡ Reduces executive reporting from hours to 30 seconds · inspired by production AI work at Workday

4

GenAI layers

30s

Time to summary

RAG

Context retrieval

OpenAI GPT-4RAGStreamlitPydantic

Medium → LinkedIn ⎆ Substack ↗ github ⎆

marketing-ai · q3 executive summary

Q3 CAMPAIGN PERFORMANCE · AI SUMMARY

4.2x

ROAS

above 3.5x target

$1.8M

Revenue

+23% vs Q2

↓18%

Email CTR

anomaly WoW

Generate Q3 executive summary.

Q3 delivered ROAS 4.2x, exceeding 3.5x benchmark by 20%. Paid search drove 62% of revenue. Email CTR declined 18% WoW — recommend creative refresh before Q4.

01 / churn

ML + AI · Flagship

Customer Churn Analysis & Prediction Copilot.

End-to-end ML system for telecom churn prediction — SHAP explainability per prediction, LLM-generated retention strategies per customer, and a FastAPI microservice. Inspired by real patterns from enterprise data work at Workday. CI/CD via GitHub Actions.

🎯 Per-customer SHAP explainability + LLM retention strategy in a single FastAPI call

SHAP

Explainability

FastAPI

Microservice

CI/CD

GitHub Actions

scikit-learnOpenAISHAP StreamlitFastAPIGitHub Actions

Medium → LinkedIn ⎆ Substack ↗ github ⎆

churn-analysis · risk dashboard

CUSTOMER RISK ANALYSIS · LIVE PREDICTION

72%

Churn score

HIGH RISK

7

Tenure (mo)

short tenure

$89

Monthly charges

above avg

Why is #CUS-4821 flagged high risk?

Month-to-month + tenure 7mo + charges above cohort median = 3 top SHAP signals. Recommend targeted retention offer within 48hrs.

04 / chatbot

AI · Deployed

GPT-4o AI Chatbot — multi-turn, deployed on Render.

Production web chatbot with multi-turn memory, animated typing indicators, and a responsive dark UI. Flask + Gunicorn backend deployed on Render. One-line system prompt swap for any domain — built as a reusable baseline for enterprise conversational AI.

🚀 Live on Render · multi-turn context window · one-line system prompt swap for any domain

GPT-4o

Model

Multi-turn

Memory

Render

Live deployed

FlaskOpenAI GPT-4oGunicornRender

Medium → LinkedIn ⎆ Substack ↗ github ⎆

ai-chatbot · multi-turn session

GPT-4o · MULTI-TURN CONVERSATION

What are the key drivers of customer churn?

Key drivers: month-to-month contracts, tenure under 12 months, charges above cohort median, lack of bundled services. Want retention strategies?

Yes, focus on retention strategies.

Top 3 levers: (1) proactive offer at 6-month mark, (2) bundle upgrade for high-usage, (3) CSM outreach when charges spike >20% MoM.

09 / kids-ai

Product · EdTech

Kids AI Explorer — AI literacy for ages 6–12.

A 4-week, 12-activity curriculum that teaches children foundational AI concepts through hands-on exploration guided by Sparky, a robot mascot. Anonymous auth on first visit with optional cross-device sync, parent coaching guides per week, real AI chat powered by Gemini or Claude, and printable completion certificates. Shipped from scratch to live on Vercel in a single sprint.

🎓 4 weeks · 12 activities · Gemini + Claude powered chat · live on Vercel

4

Weeks curriculum

12

Activities

Live

On Vercel

React 18TypeScriptTailwind CSS v4 Google GeminiSupabaseVercel

Live app ↗ github ⎆

kids-ai-explorer · week 2 · pattern recognition

WEEK 2 · PATTERN RECOGNITION · IN PROGRESS

✓Week 1 · What is AI? · 3 activities⭐⭐⭐

●Week 2 · Patterns & Predictions · 2 / 3active

○Week 3 · Train a Classifier · locked

Sparky: What comes next — 2, 4, 6, 8, ___?

10! It adds 2 each time. That's how a machine learns patterns!

✓ Badge unlocked · Pattern Spotter

Week 2 complete · 6 stars earned · Certificate unlocks after Week 4

10 / marginalia

Product · Read-it-later

Marginalia — your reading list, summarized by AI.

Paste a URL, Marginalia fetches and strips the article via Mozilla Readability, generates a TL;DR and tags via Gemini 2.5 Flash, then saves it to a personal reading list with an editorial aesthetic. Server-side API routes keep credentials off the client. Zero-cost stack: Next.js, Supabase, Stack Auth, and Vercel — all free tier, all self-hostable.

📚 URL → TL;DR in one click · $0 infra cost · editorial typography with Crimson Pro

~4s

Per summary

$0

Infra cost

SSR

Credentials safe

Next.jsTypeScriptGemini 2.5 Flash SupabaseStack AuthTailwind CSSVercel

Live app ↗ github ⎆

marginalia · reading list

READING LIST · 14 ARTICLES SAVED

https://simonwillison.net/2025/may/llm-agents

Fetching & summarizing...

◆ TL;DR · Gemini 2.5 Flash

LLMs are shifting from reactive tools to autonomous agents. Core challenges: reliable tool use and faithful long-context retrieval. #llms #agents

14

Articles saved

~4s

Per summary

$0

Infra cost

13 / kinsync

AI Agent · Family Logistics

KinSync — a family command hub with a Reasoning Engine assistant.

A scheduling hub for multi-child households — daily briefs, driver assignments, packing lists, and live Google Maps directions — backed by a conversational assistant built on Google ADK and deployed to Vertex AI Reasoning Engine. The assistant talks to the frontend over the A2A protocol using A2UI, rendering interactive Maps route cards and conflict detection directly inside the chat drawer instead of as plain text.

🗓️ Conflict-aware scheduling assistant · live Maps route cards via A2UI · Vertex AI Reasoning Engine backend

A2A

Agent-to-app protocol

A2UI 0.8

Interactive UI schema

2.5

Gemini Flash reasoning model

Google ADKVertex AI Reasoning EngineA2UI Gemini 2.5 FlashFastAPI

github ⎆

kinsync · AI assistant drawer

DAILY BRIEF · TUESDAY

✓Drop-off · Maya · soccer · Dad7:45a

✓Pickup · Leo · piano · Grandma Helen4:30p

●Conflict check · overlapping pickupsnow

How long to soccer from home?

~12 mins (3.4 mi) — route card rendered via A2UI ↓

Data-Analytics Engineering Leader /
building AI-native
data systems.

From dashboards
to decision systems.

Verified credentials.

Thirteen systems. From governed analytics to AI agents.

HR Analytics Copilot — a dashboard that writes its own SQL.

Coffee Barista Agent — an ADK agent that won't recommend off-menu.

Pipeline Sentinel — a self-healing agent for Airflow.

LoanLens — portfolio intelligence for a simulated Series C lender.

Growth Intelligence Agent — a virtual revenue analyst.

AI-Native Data Platform on Snowflake Cortex.

Real-Time Retail Sales Pipeline — Kafka to warehouse, E2E.

Marketing AI Intelligence Engine — board-ready in 30 seconds.

Customer Churn Analysis & Prediction Copilot.

GPT-4o AI Chatbot — multi-turn, deployed on Render.

Kids AI Explorer — AI literacy for ages 6–12.

Marginalia — your reading list, summarized by AI.

KinSync — a family command hub with a Reasoning Engine assistant.

The Data Brief — strategy, data, decisions.

15 years · 6 companies · one arc.

Intuitive Surgical

Workday

Lyft

Intuitive Surgical

ServiceNow

Juniper Networks

Tech Mahindra

Full lifecycle — warehouse to LLM.

Let's build something that ships.

Data-Analytics Engineering Leader / building AI-nativedata systems.

From dashboardsto decision systems.

Verified credentials.

Thirteen systems. From governed analytics to AI agents.

HR Analytics Copilot — a dashboard that writes its own SQL.

Coffee Barista Agent — an ADK agent that won't recommend off-menu.

Pipeline Sentinel — a self-healing agent for Airflow.

LoanLens — portfolio intelligence for a simulated Series C lender.

Growth Intelligence Agent — a virtual revenue analyst.

AI-Native Data Platform on Snowflake Cortex.

Real-Time Retail Sales Pipeline — Kafka to warehouse, E2E.

Marketing AI Intelligence Engine — board-ready in 30 seconds.

Customer Churn Analysis & Prediction Copilot.

GPT-4o AI Chatbot — multi-turn, deployed on Render.

Kids AI Explorer — AI literacy for ages 6–12.

Marginalia — your reading list, summarized by AI.

KinSync — a family command hub with a Reasoning Engine assistant.

The Data Brief — strategy, data, decisions.

15 years · 6 companies · one arc.

Intuitive Surgical

Workday

Lyft

Intuitive Surgical

ServiceNow

Juniper Networks

Tech Mahindra

Full lifecycle — warehouse to LLM.

Let's build something that ships.

Data-Analytics Engineering Leader /
building AI-native
data systems.

From dashboards
to decision systems.