AskAiml
Sign in Get started

Specialisation capstone — your differentiator

A polished, end-to-end project in your chosen track with evals, cost dashboards, and a public write-up.

🛠 Capstone

After Phase 6. This is the project that says "I am the candidate for this role." Pick one track from 06_specialization/01_specialization_tracks.md and ship a polished demo.

The template below works for any track — fill in the blanks.

1. One-pager spec

Copy this template into your repo as SPEC.md before writing code.

# Project: <name>
Owner: <you>
Track: voice / browser / coding / multimodal / health / finance / robotics / security
Date: <YYYY-MM-DD>

## Problem
<who hurts, why now, what does success look like?>

## Target user
<persona, scenario, channel>

## Success metric
<single primary metric, with target>

## Anti-goals
<things this project will NOT do>

## Constraints
<latency, cost, privacy, regulatory>

## High-level architecture
<mermaid diagram>

## Tech choices
<bullet list with reasons>

## Eval plan
<datasets, metrics, gates>

## Deployment plan
<where it runs, how it scales, observability>

## Risks & mitigations
<top 5>

A spec like this is exactly what an L5/L6 engineer would write before shipping. Hiring managers read it first.

2. Architecture template (mermaid)

Adapt to your track. Examples:

Voice agent

mermaid
flowchart LR
  PHN[Phone / Web call] --> RT[LiveKit / Twilio Media]
  RT --> STT[Whisper Large v4]
  STT --> AGENT[LangGraph agent]
  AGENT --> TOOLS[KB / CRM / calendar tools via MCP]
  AGENT --> TTS[ElevenLabs v3]
  TTS --> RT
  AGENT --> OBS[LangSmith + Grafana]

Browser agent

mermaid
flowchart LR
  USER --> UI[Streamlit/CLI]
  UI --> ORCH[LangGraph orchestrator]
  ORCH --> CU[Claude Computer Use]
  CU --> BB[Browserbase]
  BB --> SITE[Target site]
  ORCH --> HIL[interrupt() on unknown screen]
  ORCH --> OBS[LangSmith]

Coding agent

mermaid
flowchart LR
  PR[GitHub PR] -->|webhook| GH[GitHub MCP server]
  GH --> AGENT[LangGraph reviewer]
  AGENT --> RG[ripgrep / tree-sitter]
  AGENT --> TESTS[pytest sandbox]
  AGENT --> POST[Inline PR comments]

3. The 4-week plan

Week 1 — MVP thin slice

  • Spec + architecture diagram approved by yourself or a peer.
  • End-to-end ugly-but-working flow on a single happy-path input.
  • Push to GitHub day 1; commit daily.

Week 2 — Quality

  • Build a 30-50 case eval set covering happy / edge / adversarial.
  • Add observability (LangSmith + Grafana).
  • Tune prompts / retrieval / model choice based on eval.

Week 3 — Production polish

  • Guardrails, retries, circuit breakers.
  • Cost meter and dashboard.
  • HIL or refusal flow where appropriate.
  • Containerise, deploy to a cheap cloud VM or Cloud Run.

Week 4 — Story

  • Write the post (1500-2000 words).
  • Record 60-90 second video demo.
  • Optional: add a public chat / try-it-yourself link with rate-limit.
  • Apply for jobs. Reference this project in cover letters.

4. Story template (the thing recruiters actually read)

Sections:

  1. What it does (1 paragraph + screenshot/gif).
  2. Why it matters (the problem and stakes).
  3. Architecture (mermaid + 1 paragraph per box).
  4. The 3 hardest decisions and why (e.g., "we chose pgvector over Qdrant because...").
  5. Evaluation results (table with baseline vs final).
  6. Cost / latency (numbers).
  7. What I would do next (honesty wins).
  8. Code links (repo, key files, demo).

If a recruiter only reads 3 paragraphs, sections 1, 4, 5 should be those.

5. Checklist before applying

  • Public GitHub repo with clear README.
  • LICENCE file (MIT or Apache 2.0).
  • CI green (lint + tests + eval gate).
  • Docker / Compose runs locally with one command.
  • Live demo URL or Loom video.
  • Mermaid architecture diagram.
  • Eval table with numbers.
  • Cost analysis.
  • At least one mention of guardrails / safety.
  • Honest "future work" section.

When you can tick all boxes — start applying. You are ready.