Specialisation tracks — pick one and ship the capstone

Voice, browser, coding, multimodal, and domain tracks, with the signature capstone shape for each.

🎯 Module 6 6 min read Not started

Why this matters

After Phases 0-5 you can ship general-purpose agents. To stand out (and to command the higher salary bands), pick a specialisation. Hiring managers grade for breadth and one deep area. This lesson outlines the 8 highest-leverage specialisations in 2026 and how to enter each.

Pick one. Ship one capstone in it. The rest stay on your radar.


1. Voice agents

What it is: real-time speech-in / speech-out agents that hold conversations.

Stack:

  • Speech-to-text: OpenAI Whisper Large v4, Deepgram Nova-3, Speechmatics, AWS Transcribe.
  • Realtime API: OpenAI Realtime API (websocket), Anthropic streaming, Hume AI.
  • Telephony: Twilio Voice + Media Streams, LiveKit, Vapi, Retell AI.
  • TTS: ElevenLabs v3, OpenAI TTS, Cartesia Sonic, Google Cloud TTS.
  • Latency tricks: voice activity detection, partial-token streaming, semantic interruption.

Capstone idea: a Hindi/English bilingual receptionist agent for a clinic that books appointments, answers FAQs, and escalates to a human after 3 failed attempts. Sub-2-second turn latency.

Where the jobs are: customer support, healthcare scheduling, sales, automotive, banking IVR.


2. Browser / computer-use agents

What it is: agents that drive a real browser or desktop to do tasks (form fills, scraping, ops, QA).

Stack:

  • Browser: Playwright (Python), Puppeteer (Node), Browserbase (managed).
  • Vision: Claude Computer Use, OpenAI computer-use tool, Anthropic Skills.
  • Wrappers: browser-use (community), AutoGen browser tools, Skyvern.
  • Sandboxing: E2B, Modal sandboxes, Daytona.

Capstone idea: an agent that automates expense-report filing on your company portal — login, fill, attach, submit, screenshot proof. Failure mode: pause for human at any unexpected screen.

Beware: legal grey areas (CAPTCHAs, ToS). Use only with explicit consent or your own systems.


3. Coding / SWE agents

What it is: agents that read repos, write code, run tests, open PRs.

Stack:

  • Models: GPT-5-codex, Claude Sonnet 4.6 (very strong on SWE-bench), DeepSeek-Coder, Qwen2.5-Coder.
  • Tool surfaces: ripgrep, AST tools (tree-sitter), file editing, shell, MCP servers for Git/GitHub.
  • Agents: SWE-agent, OpenDevin, Aider, Cursor Agents, Devin.
  • Benchmarks: SWE-bench Verified, LiveCodeBench, MBXP, BigCodeBench.

Capstone idea: a "code reviewer" agent that, given a PR diff, runs static analysis, references your style guide, runs tests, and posts inline GitHub PR comments via the GitHub MCP server.


4. Multimodal / vision agents

What it is: agents that see (images, video, screens, diagrams).

Stack:

  • Models: GPT-4o, Claude Sonnet 4.6 vision, Gemini 2.5 Pro, Qwen2.5-VL.
  • Document AI: LlamaParse, Docling, Unstructured, Mathpix, Azure Document Intelligence.
  • Charts: code-interpreter / matplotlib for generation; OCR + vision for parsing.

Capstone idea: an "expense-receipt agent" that reads a photo of any receipt (English/regional Indian languages), extracts items + GST + totals into Pydantic, books the entry into Tally / Zoho Books.


5. Healthcare and life sciences

What it is: agents that assist clinicians, summarise records, draft notes, do triage.

Stack:

  • De-identification: Presidio + medical entity rules, OHDSI, NLM Scrubber.
  • Knowledge sources: PubMed, UpToDate, ClinicalKey, FDA labels.
  • Domain models: OpenBio-LLM, Med-PaLM 2, MedGemma.
  • Compliance: HIPAA (US), DPDP (India), GDPR (EU). On-prem deployment is common.

Capstone idea: a SOAP-note assistant that listens to a doctor-patient conversation (recorded), drafts the SOAP note, flags red-flags, and never stores PHI outside the clinic's VPC.

Salary premium: high. Risk: high. Process: clinical validation, not just engineering eval.


What it is: agents that read 10-Ks, contracts, regulatory filings; draft memos; perform due diligence.

Stack:

  • Document parsing: LlamaParse for tables and signatures, Docling, AWS Textract.
  • Reasoning models: Claude Opus 4.7, GPT-5.5 (long-context tasks).
  • Knowledge graphs: Neo4j or Memgraph for entity relationships.
  • Audit: every claim must have a citation; retrieval must be deterministic-replayable.

Capstone idea: a contract-risk reviewer that takes a 60-page MSA and produces a markdown table of clauses, highlighting deviations from your company's standard playbook.


7. Robotics / embodied agents

What it is: agents that perceive and act in the physical world (robots, drones, factory).

Stack:

  • Perception: YOLO-world, SAM-2, depth (RealSense, ZED), point-cloud (Open3D).
  • Planning: OpenVLA, π0 (Physical Intelligence), Octo, RT-2.
  • Sim: NVIDIA Isaac Sim / Gym, MuJoCo, Genesis.
  • Hardware: ROS 2, MoveIt 2, Foxglove.

Capstone idea: simulate a warehouse pick-and-place policy in Isaac Sim, evaluate, then deploy to a low-cost arm. Heavy ML / control theory; fewer LLM hallucinations, more "did the model crash the gripper?"

Hardest path. Highest pay if you nail it.


8. AI security / red-teaming

What it is: attacking and defending LLM systems.

Stack:

  • Attack tools: Garak, PyRIT (Microsoft), Promptfoo, Lakera Guard tests.
  • Bench: AgentBench, ASB, AdvBench, JBB.
  • Defenses: NeMo, Presidio, Llama Guard, Lakera, Rebuff.
  • Frameworks: OWASP LLM Top 10, MITRE ATLAS, NIST AI RMF.

Capstone idea: red-team your own agent (Phase 5 stack) with Garak + PyRIT; document every successful exploit; close them; post the case study.


How to choose

Pick the track that fits your background:

  • Backend → coding agents, SWE.
  • Frontend / UX → voice or browser agents.
  • Domain expert (doctor, lawyer, accountant) → that vertical.
  • Curious / breadth-first → AI security / red-teaming.

Start the capstone in week 1 of this phase. Aim for a polished demo + write-up by week 4.


Capstone playbook (5-step)

  1. Scope on a single page. Problem, target user, success metric, anti-goals.
  2. Architecture diagram. Mermaid; label every box and the data flow.
  3. MVP build (1-2 weeks). End-to-end thin slice. Ugly UI is fine.
  4. Evaluation set + dashboard. Track quality, cost, latency.
  5. Polish + write-up. Public repo, live demo, 1500-word post explaining decisions.

A capstone is judged by the post, not the code. Spend a full day on the write-up. Include screenshots, mermaid diagrams, eval tables, cost analysis, and lessons learned.


Reference job titles per track

TrackCommon titles
VoiceVoice AI Engineer, Conversational AI Engineer
Browser/computer-useAutomation Engineer, Browser Agent Engineer
CodingAI SWE, Coding Agent Engineer, DevAI
MultimodalVision AI Engineer, Document AI Engineer
HealthcareClinical AI Engineer, Health AI ML
Finance/legalLegalTech AI, FinTech AI Engineer
RoboticsML Robotics, Foundation-Model Robotics
SecurityLLM Red Team, AI Security Engineer

Salary premiums in 2026 (roughly): security and healthcare top, then voice and finance, then SWE, then everything else.


Final words

The curriculum has taken you from print("hello") to "I run a multi-agent, evaluated, monitored, possibly self-hosted, possibly fine-tuned agent product." That is a senior engineer's resume. Pick a specialisation, ship the capstone, polish the README, and start applying.

The agent industry in 2026 needs people who can ship — not people who have read about it. Ship the project.


References

Sign in to save your progress and earn badges.