Class 12 · 12.2

FiNALLY project: aapka multi-agent capstone

Is haftay ki har technique ko rehne ke liye aik jagah chahiye. Woh jagah FiNALLY hai — “Finance Ally” — aik shandaar AI-powered trading workstation jo hum agents ki fauj ke sath banayenge. Yahan hum woh bunyaad rakhte hain jo controlled chaos ko mumkin banati hai.

🎯 Maqsad: bunyaad rakhna⏱ Parhne ka waqt: 14 min🔧 Qism: project setup

FiNALLY naam ke mutabiq aik capstone hai: yeh AI co-pilot wale modern Bloomberg terminal jaisa nazar aata aur mehsoos hota hai. Agar finance aapke liye nayi hai to fikr na karein — agar stocks aur shares ka pata hai to aap kafi jaante hain.

01Vision

Product live market data stream karta hai, users ko simulated portfolio trade karne deta hai, aur aik LLM assistant shamil karta hai jo portfolio analyse kar sakta hai aur aapki taraf se trades bhi execute kar sakta hai. Yeh course ka capstone hai, jo poori tarah coding agents ne banaya — is baat ka saboot ke orchestrated agents production-quality, full-stack app bana sakte hain.

Bunyadi architectural choice

Agents planning/ directory ki files ke zariye interact karte hain. Yeh aik faisla — aik shared jagah jahan woh aik doosre ke liye notes chhor saken — independent agents ke dher ko coordinated team mein badalta hai. 12.1 ka “files as shared memory” yaad hai? Amal mein yahi hai.

02Scaffolding clone karein

Aap aik chhoti, lagbhag khaali repo se shuru karte hain — aik skeleton jise agents mukammal karenge.

terminal

# into your projects directory
cd ~/projects
git clone <the-finally-starter-repo> finally
cd finally
code .   # open it in VS Code

Andar starting point jaan boojh kar saada rakha gaya hai:

finally/

README.md          # nearly empty
CLAUDE.md          # points everything at planning/
.env               # your OpenRouter key
.gitignore
planning/
  plan.md          # the business requirements — the heart of it
backend/           # empty
frontend/          # empty
db/                # empty
tests/             # empty
.claude/skills/    # the Cerebras skill, copied from week 2

Purane dost wapas

Ghaur karein ke .claude/skills/ folder mein week two mein likhi hui Cerebras skill pehle se mojood hai. Folder rakh dein aur skill foran kaam karti hai — koi setup nahi. Skills ki khoobsurti yahi hai, aur isi liye baad mein unhein plugins mein bundle karte hain.

03plan.md — aik source of truth

Poora build planning/plan.md ke gird ghoomta hai. CLAUDE.md chhoti hai aur aik kaam karti hai: @ notation se har baar poora plan context mein laati hai, taake koi agent use miss na kar sake.

CLAUDE.md

# FiNALLY — the Finance Ally

All project documentation lives in the planning directory.
The key document is the plan, included in full below:

@planning/plan.md

@ kyun aham hai

@planning/plan.md reference har turn par plan ko context mein laane par majboor karti hai — iske paas koi choice nahi. Is tarah plan.md aik hi, hamesha mojood source of truth ban jati hai: business requirements, architecture, aur rules. Plan ko control karein aur build ko control karein.

Is plan ko likhne ka acha tareeqa (aur yahan bhi yahi kiya gaya): vision ka aik imaandaar paragraph likhein, use Claude chat product mein paste karein, aur question-and-answer ke zariye tab tak iterate karein jab tak requirements mazboot mehsoos na hon. Human judgement rehnumai karta hai; model draft banata hai.

04Parallel work ke liye boundaries

Agar kai agents aik waqt mein mukhtalif hisse banayenge to plan ka sab se aham section woh hai jo boundaries bayan karta hai — kis ki zimmedari kya hai aur pieces aik doosre se kaise baat karte hain.

Har agent aik area ka malik hota hai aur plan ke zariye coordinate karta hai. Saaf boundaries hi parallel agents ko aik doosre ke kaam mein dakhal dene se rokti hain.

Swarm bohot jaldi launch na karein

Jab plan abhi kamzor ho to teen agents (“frontend build karo”, “backend build karo”, “tests build karo”) shuru karna aksar fiasco par khatam hoga — aise pieces jo fit nahi hote aur aise agents jo aik doosre se be-khabar hain. Pehle mazboot bunyaad banayein; phir parallelise karein. Yahi controlled chaos hai.

05Architecture, jaan boojh kar saada

Plan stack tay karta hai aur insaan ise lean rakhta hai:

Concern	Choice	Kyun
Market data	Default par simulated; optional API key se real	Aik paisa diye baghair live mehsoos hota hai; baad mein real data laga sakte hain.
Streaming	SSE (server-sent events)	Streaming LLM output jaisa hi pattern — prices ko UI par live push karta hai.
Database	SQLite, sessions ke darmiyan persist	Abhi saada, baad mein Postgres/Supabase par upgrade karna aasaan.
LLM	OpenRouter / Cerebras + structured outputs	Tez hai aur assistant ko trade decisions bharose ke sath lene deta hai.
Deploy	Aik Docker container	LLMs kai containers ke sath over-engineer karna pasand karte hain — insaan ne kaha “nahi, aik hi rakho.”

Insaan ka sab se qeemti qadam

Jab model ne multiple Docker containers ki taraf rukh kiya to instructor ne sawal uthaya: kya yeh aur saada ho sakta hai? Ise aik hi container tak rakhna aisa seedha, simplifying sawal hai jo sirf aap pooch sakte hain — aur yahin aap sab se zyada value add karte hain.

✓ Ahm baatein

FiNALLY week ka capstone hai: agents ka banaya hua live AI trading workstation.
Agents planning/ ki files ke zariye coordinate karte hain — shared memory amal mein.
plan.md single source of truth hai; @plan.md har turn par ise context mein laati hai.
Saaf boundaries agents ko takraye baghair parallel kaam karne deti hain — lekin pehle bunyaad banayein.
Architecture simple rakhein (aik Docker container, SQLite, SSE) — over-engineering ko challenge karein.