Agentic AI Systems Β Β·Β AI/ML Engineering Β Β·Β RAG & LLM Applications Β Β·Β Python Backends Β Β·Β Data Engineering
I'm an Agentic AI & Systems Engineer who likes owning a problem end-to-end, from the data and the retrieval layer through the model to the product people actually use.
- π I build autonomous Python pipelines, PySpark / Databricks lakehouses, and RAG / LLM systems, then ship them as full-stack apps with Next.js & FastAPI.
- π§© Deepest in agents, RAG and Python backends, with real range across data engineering and full-stack when the problem needs it.
- π B.Tech CSE, Class of 2026 (graduated). Earned an M.Tech CSE seat at IIIT Tiruchirappalli and turned it down to build full-time.
- β‘ I default to free, production-grade tooling: most of my projects run at ~$0/month.
- π« Open to AI / ML / Backend roles, right now. Let's talk: shaansatsangi@gmail.com
The same spine runs through almost everything below: the model writes the prose, never the number. Decisions live in code that can be tested; the LLM narrates, routes, and asks for clarification around it. Every layer fails open, except the gates, which fail closed.
flowchart LR
IN["<b>Messy input</b><br/>APIs Β· webhooks<br/>PDFs Β· voice Β· sheets<br/>event streams"]
subgraph CORE ["Deterministic core: the part that must be right"]
direction TB
ING["<b>Typed ingestion</b><br/>rate-limit aware<br/>conditional requests<br/>bounded fan-out"]
LOGIC["<b>Pure decision code</b><br/>scoring Β· matching Β· salience<br/><i>every output traces to evidence</i>"]
CACHE["<b>Fail-open cache</b><br/>Redis Β· Postgres Β· Parquet"]
ING --> LOGIC --> CACHE
end
subgraph EDGE ["LLM at the edges"]
direction TB
NARR["<b>Narration & synthesis</b><br/>receives computed JSON,<br/>never the raw data"]
ROUTE["<b>Tool routing</b><br/>closure-bound per user<br/>failover across model chains"]
end
GATE{"<b>Gate</b><br/>is this worth<br/>sending at all?"}
OUT["<b>Shipped surface</b><br/>Next.js Β· FastAPI<br/>Telegram Β· MCP Β· CLI"]
SILENT(["<b>Silence</b><br/><i>audited, never accidental</i>"])
IN --> ING
LOGIC --> NARR
LOGIC --> ROUTE
NARR --> GATE
ROUTE --> GATE
CACHE --> OUT
GATE -->|"clears the bar"| OUT
GATE -->|"doesn't, or fails"| SILENT
classDef core fill:#1e293b,stroke:#a78bfa,stroke-width:2px,color:#e2e8f0
classDef edge fill:#312e81,stroke:#818cf8,stroke-width:2px,color:#e2e8f0
classDef term fill:#0f172a,stroke:#475569,stroke-width:1.5px,color:#cbd5e1
class ING,LOGIC,CACHE core
class NARR,ROUTE edge
class IN,OUT,SILENT,GATE term
Where each piece is real, not aspirational: the deterministic core is Skill Issue's two-pass scorer Β· the gate is Atlas's salience check, which returns send: false and means it Β· closure-bound routing is Atlas's 20 tools, where every tool that touches user data is welded to one user id Β· fail-open caching is Skill Issue's four Redis layers.
flowchart TD
ME(("<b>Shaan</b>"))
ME --> A["<b>Agentic AI</b>"]
ME --> B["<b>RAG & retrieval</b>"]
ME --> C["<b>Python backends</b>"]
ME --> D["<b>Data engineering</b>"]
A --> A1["<b>Atlas</b>: 20 tools under Gemini<br/>function calling, live on Telegram"]
A --> A2["<b>Syntrueno</b>: agent swarm that mutates<br/>real Cloud Run behind a human gate"]
A --> A3["<b>telegram-mcp</b>: MCP server giving<br/>agents your own Telegram account"]
B --> B1["<b>Think9 Brain</b>: cites its source section,<br/>refuses when the corpus can't answer"]
B --> B2["<b>disaster-management-rag</b>: 18/20 held-out,<br/>reproducible with no API key"]
B --> B3["<b>JARVIS-PY</b>: semantic memory +<br/>threshold-gated PDF RAG, local-first"]
C --> C1["<b>Skill Issue</b>: FastAPI, Neon, Upstash,<br/>OAuth, four-layer Redis cache"]
C --> C2["<b>VouchRank</b>: multi-tenant SaaS on<br/>Postgres RLS + 12 Deno edge functions"]
D --> D1["<b>YouTube Wrapped</b>: Databricks medallion<br/>lakehouse on Delta Lake"]
D --> D2["<b>CRM Sales Warehouse</b>: Airflow-orchestrated<br/>star schema, SQL quality checks, Power BI"]
classDef root fill:#a78bfa,stroke:#7c3aed,stroke-width:3px,color:#0f172a
classDef pillar fill:#312e81,stroke:#818cf8,stroke-width:2px,color:#e2e8f0
classDef proof fill:#1e293b,stroke:#475569,stroke-width:1.5px,color:#cbd5e1
class ME root
class A,B,C,D pillar
class A1,A2,A3,B1,B2,B3,C1,C2,D1,D2 proof
π Atlas Β Β·Β βΆ Talk to it
An AI financial analyst that lives in Telegram: conversation only, no slash commands or buttons. Learns you as you talk, pulls live quotes and fundamentals through provider failover chains (four quote, three fundamentals), reads PDFs and spreadsheets natively (tables stay tables), transcribes voice notes, and sends a morning briefing only when something on your watchlist matters; silence is enforced control flow, not a prompt suggestion. Gemini function calling drives the loop across 20 tools; the ten that touch user data are closure-bound to one user, so the model cannot reach another's data. Always-on under |
β‘ Skill Issue
Reads any public GitHub profile and scores it out of 100 across six weighted signals, then places you on a 7-tier ladder from Hobbyist to Principal with a sub-rank. Every point traces to real evidence in your repos; the AI writes the narrative, never the numbers (0 hallucinated scores). Pick Roast Mode for what a blunt senior would say, or Mentor Mode for how to fix it. Free, no signup; sign in to save reports and share them.
|
π¬ VouchRank Β Β·Β βΆ Live Demo
Multi-tenant, white-label reputation + AI-search-optimization SaaS: a compliance-first review funnel (FTC / Google 2026 review rules), LLM-visibility audits (ChatGPT Β· Gemini Β· Perplexity), embeddable social-proof widgets, Stripe subscriptions + billing portal, and Supabase Postgres Row-Level Security tenant isolation, plus 12 Deno edge functions that check agency ownership before touching data. Fully clickable live in demo mode. |
π¬ AI Reel Factory
On-demand, low-cost pipeline (a $5/month target) that researches news, writes & narrates scripts in a near-human voice, generates AI B-roll, auto-edits a captioned cinematic vertical video, and uploads it as a YouTube Short. A |
π YouTube Wrapped
Personal watch-history analytics product built on a Databricks medallion lakehouse (Bronze β Silver β Gold) over Delta Lake, served through a deployed FastAPI + Next.js dashboard. |
ποΈ CRM Sales Warehouse
End-to-end CRM analytics platform: Python ETL/ELT into a star-schema PostgreSQL warehouse, orchestrated with Airflow, guarded by SQL data-quality checks plus a starter dbt layer with its own tests, and surfaced as a 5-page Power BI executive dashboard. |
π§ JARVIS-PY
Local-first AI voice assistant with wake-word barge-in, online/offline STT (Google, or Vosk on-device), semantic memory, PDF RAG, tool-agent routing, reminders, and interruptible TTS. The LLM runs locally through Ollama, and it keeps working offline. |
π‘οΈ Sahaara
Safety-focused full-stack app: a one-tap emergency SOS broadcasts an alert to online users within a set radius over Supabase Realtime, with MapLibre live location tracking and a trusted-contacts list. |
π°οΈ Syntrueno Β Β·Β βΆ Live service
A zero-trust autonomous cloud-ops swarm. Gemini agents diagnose live incidents, propose remediations and judge their own plans for safety. Behind a single-use SHA-256 human gate bound to that exact action, they execute real changes against a real Cloud Run service, then re-read live state until it converges rather than trusting the API's acknowledgement. Guards fail closed, cheapest first. Every guard refusal is written to a hash-chained Firestore ledger, never dropped silently. 288 tests run offline in ~2.7 s with no credentials. |
π§Ύ Think9 Brain Β Β·Β βΆ Live
A grounded institutional-memory assistant over a synthetic multi-brand corpus. It cites the exact source section, states a good-as-of date, and holds back superseded documents. Built to be trusted, which means built to say "I don't know." Refusal recall is 1.000 on both eval sets: across 32 unanswerable questions (invented vendors, plausible-but-absent figures, out-of-scope functions) it never once fabricated an answer. |
More: telegram-mcp (MCP server for your own Telegram account) Β· Face Sort Studio (local CV photo organizer) Β· Cog & Cosmos (Svelte/PixiJS idle game) Β· and browse all repos β
Shipping into other people's codebases, where the review bar isn't mine to set.
Merged upstream
| PR | Project | What it fixed |
|---|---|---|
| ha-mcp#1670 | Home Assistant MCP Β· 5.0k β | A restart through a reverse proxy returned 502/503 and was read as a failure; the known-good patterns now cover it |
| android#7018 | Home Assistant Android Β· 4.0k β | Refactored the media-player widget config activity to Compose + Material 3 |
| narwhals#3697 | Narwhals Β· 1.7k β | Added a maintain_order argument to Expr.list.unique / Series.list.unique, across the backend implementations |
| anker-solix-api#314 | anker-solix-api Β· 304 β | Extended X1 device support by updating the status enumerations |
In review: ollama#16715 (Ollama, 183k β: CLI image-path parsing) Β· supabase#46915 (Supabase, 111k β: newlines in the Studio SMS OTP template) Β· helm#32250 (Helm, 30k β: show the installed chart's source) Β· dbt#12952 (dbt Core, 14k β: clearer snapshot unique_key errors) Β· langchain-google#1950 and #1840 (LangChain Google, 404 β: no spurious $defs warning, async integration tests)
flowchart LR
ME(("<b>Contributions</b>"))
ME --> DATA["<b>Data & DataFrames</b>"]
ME --> INFRA["<b>Infrastructure</b>"]
ME --> AI["<b>AI tooling</b>"]
ME --> IOT["<b>Home & IoT</b>"]
DATA --> N["Narwhals<br/><b>merged β
</b>"]
DATA --> DBT["dbt Core<br/><i>in review</i>"]
INFRA --> H["Helm<br/><i>in review</i>"]
INFRA --> S["Supabase<br/><i>in review</i>"]
AI --> O["Ollama<br/><i>in review</i>"]
AI --> LG["LangChain Google<br/><i>in review Γ2</i>"]
IOT --> HM["Home Assistant MCP<br/><b>merged β
</b>"]
IOT --> HA["Home Assistant Android<br/><b>merged β
</b>"]
IOT --> AS["anker-solix-api<br/><b>merged β
</b>"]
classDef root fill:#a78bfa,stroke:#7c3aed,stroke-width:3px,color:#0f172a
classDef area fill:#312e81,stroke:#818cf8,stroke-width:2px,color:#e2e8f0
classDef merged fill:#064e3b,stroke:#10b981,stroke-width:2px,color:#d1fae5
classDef review fill:#1e293b,stroke:#475569,stroke-width:1.5px,color:#cbd5e1
class ME root
class DATA,INFRA,AI,IOT area
class N,HM,HA,AS merged
class H,S,O,LG,DBT review
ποΈ Data Engineering & Analytics
π€ AI / ML
π» Full-Stack
βοΈ DevOps & Cloud
+ more tools I reach for
Also familiar with, though not yet in a public repo: TensorFlow, Keras, Tableau, Seaborn, NLTK, Twilio, C++.
- π°οΈ Built Syntrueno for the Google Cloud "All Things Agentic" Hackathon: an agent swarm that makes real infrastructure changes behind a cryptographic human gate. Live on Cloud Run.
- π Shipping Skill Issue at skillissue.tech, currently v1.0.13, which closed two critical Next.js security advisories. Scores any GitHub profile out of 100, roasts it, and hands you the receipt. Free, no signup.
- ποΈ Shipped JARVIS v3.6.0 (Sep 2026): deterministic answers for time, date, battery and disk, an offline calculator that understands spoken numbers, a hard two-sentence reply cap, and Stop/Esc that halts a streaming answer.
- π Running Atlas live on Telegram (@AtlasAnalyst_bot); always-on AI financial analyst, self-healing, 309 tests. Earned a Certificate of Distinction at the Atlas AI Hackathon (Humanity Founders).
- π¬ Running AI Reel Factory: an on-demand, low-cost pipeline that pitches news stories to me on Telegram, then scripts, voices, captions and publishes the ones I approve as Shorts on @butitmatters, 80+ so far.
- π Contributing upstream: merged into Home Assistant MCP, Home Assistant Android, Narwhals, and anker-solix-api; open PRs at Ollama, Supabase, Helm, dbt Core, and LangChain Google.
- π€ Open to AI/ML Β· Agentic AI Β· Backend Β· Data Engineering roles and available immediately.
| Certification | Issuer | Year |
|---|---|---|
| π Certificate of Distinction: Atlas AI Hackathon | Humanity Founders | 2026 |
| π M.Tech CSE admission offer (declined for full-time work) | IIIT Tiruchirappalli | 2026 |
| Introduction to Model Context Protocol (MCP) | Anthropic | 2026 |
| Data Science, AI/ML Engineer & Data Engineer | micro1 | 2026 |
| Fundamentals of Machine Learning | Microsoft | 2023 |
| Fundamental AI Concepts | Microsoft | 2023 |
| π₯ 2nd Place: IEEE Debate Competition | IEEE | - |
- π Pushed 1 commit to Shaan-alpha/Shaan-alpha (Oct 12, 2026)
- π Pushed 1 commit to Shaan-alpha/defaultradar-msme (Oct 12, 2026)
- π Pushed 1 commit to Shaan-alpha/Portfolio (Oct 12, 2026)
- π Pushed 1 commit to Shaan-alpha/Skill-Issue (Oct 12, 2026)
- β‘ Opened PR "Correct README claims and clear new Next.js advisories" in Shaan-alpha/Skill-Issue (Oct 12, 2026)
| Visitor | Message | Date |
|---|---|---|
| π Want to appear here? Sign the guestbook above! | ||
I like turning messy inputs into useful outcomes: data platforms, AI systems, and products that ship.
Always happy to talk data, AI, and building things.



