Purpose of this document. Before writing a single line of code, we convened a council of five expert personas to decide what to build, in which order, for whom, and how to sell it. Everything else in this repository (guides, code, learning material, book) follows the decisions taken here.
The council members
| 1 |
Dr. Ayesha — the Research Scientist |
12 years in plant genomics & bioinformatics, supervises 6 PhD students, publishes in Q1 journals |
Scientific rigour, real researcher pain points, reproducibility, ethics of AI in research |
| 2 |
Bilal — the AI Agent Architect |
Built production LLM systems (RAG, multi-agent orchestration, tool use, evals) |
Architecture, tool design, reliability, cost, latency, evaluation |
| 3 |
Sara — the Product & Go-to-Market Lead |
Ex-EdTech/SaaS founder, sold to universities in Pakistan, GCC and UK |
Pricing, positioning, procurement cycles, who signs the cheque |
| 4 |
Prof. Khan — the University Administrator |
Dean of Research at a public-sector university, sits on HEC/ORIC committees |
Institutional buying, compliance, data governance, training needs |
| 5 |
Hamza — the Python Educator (Codanics voice) |
Teaches “Python ka Chilla: Build AI Agents”; thousands of students |
Teachability, clean code, step-by-step learning path, community building |
The question put to the council
“We have five candidate AI research agents (Discovery, Literature Review, Experiment, Writing & Publication, PhD Supervisor) and a vision of an AI Research OS. What do we build first, how do we build it, how do we sell it, and how do we teach it?”
Round 1 — Each member’s opening position
Dr. Ayesha (Research Scientist)
- The Literature Review Agent solves the most painful, most measurable problem: a systematic review is 6–12 months of labour. If we cut it to 2–4 weeks with an auditable PRISMA trail, every graduate school will want it.
- Warning: researchers will not trust a black box. Every claim must be traceable to a real paper (DOI/arXiv ID). Hallucinated citations = instant loss of trust.
- The Experiment Agent is high value but domain-specific (RNA-seq ≠ soil science ≠ NLP). Build it as a framework with pluggable pipelines, not a one-size-fits-all.
Bilal (AI Agent Architect)
- All five agents share ~70% of infrastructure: an LLM client, paper-search tools (arXiv, PubMed, Semantic Scholar, Crossref), structured-output extraction, a planner → worker → critic loop, and a report writer.
- Recommendation: build each agent as an independent, copy-able folder (students must be able to run one agent without the others), but follow the same internal pattern so learning transfers.
- Use structured outputs (Pydantic) everywhere we need tables (screening decisions, extraction tables, reviewer comments). Free-text for prose sections only.
- Insist on evals from day one: a small “golden set” of topics with expected papers/gaps.
Sara (Product & GTM)
- Individual researchers pay $20–50/month; institutions pay $20k–$500k/year. Sell to institutions, market to individuals. Individuals become the champions who push procurement.
- Sequence: Literature Review Agent (fastest “wow”, clear ROI) → Discovery Agent (top-of-funnel, viral demo) → Writing & Publication Agent (recurring revenue: every paper) → PhD Supervisor (retention) → Experiment Agent (enterprise/pharma upsell).
- Emerging markets (Pakistan, GCC, South/South-East Asia, Africa) are underserved. Pricing must be tiered by region; HEC-affiliated universities in Pakistan can be reached via ORICs (Offices of Research, Innovation & Commercialization).
Prof. Khan (University Administrator)
- Universities buy on: (1) plagiarism/ethics safety, (2) data privacy (student data stays in-country or on-prem), (3) training & support, (4) a report we can show to accreditation bodies.
- A PhD Supervisor Agent would be politically sensitive if it replaces supervisors — position it as a supervisor’s assistant and a student’s study coach.
- Procurement cycle is 3–9 months. Offer a free pilot for one department with a success metric (e.g., “reduce time-to-literature-review by 60%”).
Hamza (Educator, Codanics)
- Students must be able to run every agent with one API key, one
.venv, and one python main.py command.
- Each agent should teach one new concept: tools & search APIs (Discovery), structured outputs & pipelines (Lit Review), code + data + stats (Experiment), multi-persona simulation (Writing), memory & long-running state (Supervisor).
- All materials must be branded and free to learn (YouTube playlist + WhatsApp channel), with the paid course as the deep-dive. Everything ends up in a Quarto book so it doubles as course material.
Round 2 — Debate & resolutions
| “Build the whole Research OS now” vs. “ship one agent” |
Ship the five agents as independent modules first, with a thin orchestrator (AI Research OS) described in the guides as Phase 2. Independent modules are easier to sell, teach and test. |
| Free-text LLM output vs. structured data |
Structured outputs (Pydantic) for every decision/table; free text only for narrative sections. This is non-negotiable for trust and for building a UI later. |
| Which LLM / provider |
Default to the Anthropic Claude API (claude-opus-5, adaptive thinking, streaming). Keep a single llm.py per agent so swapping providers is a one-file change. |
| Real search APIs vs. mock data |
Real, free APIs (arXiv, Semantic Scholar, PubMed E-utilities, Crossref) with graceful offline fallbacks so students without internet access can still run demos. |
| Hallucinated citations |
Every citation must originate from a tool result, never from the model’s memory. The writer agents receive only the retrieved paper list and are told to cite only from it. |
| Supervisor agent ethics |
Reframe as “Virtual Research Mentor” — assistant to supervisors, coach to students. Include an explicit “AI limitations & academic integrity” notice in outputs. |
| Experiment agent scope |
Ship a general tabular/omics-style pipeline (QC → differential analysis → ML → stats → methods text) using synthetic demo data, designed to be extended with domain plugins. |
The decision
What we build (in this repository)
- Five independent Python agents in
02_agents_code/, each with its own README.md, main.py, agents.py, tools.py, llm.py and outputs/.
- One guide per agent in
01_guides/ covering: problem, solution, architecture, workflow, build steps, market value, pricing, how to sell, risks, roadmap.
- A learning pack in
03_learning/: what to learn, glossary of major terms, how the agents work (patterns), resources.
- A Quarto book in
04_quarto_book/ that stitches everything into a publishable book.
Build order (product roadmap)
| 0 |
0–1 |
This repo: 5 CLI agents + guides + book |
Runs end-to-end on one API key |
| 1 |
1–3 |
Literature Review Agent as a web app (Streamlit/FastAPI) + PDF ingestion |
10 pilot users, 60% time saved |
| 2 |
3–6 |
Discovery + Writing agents in the same app; user accounts; export to Word/LaTeX |
100 paying individuals |
| 3 |
6–12 |
Institutional edition: SSO, on-prem/VPC option, admin dashboard, usage reports |
2 university licences |
| 4 |
12–18 |
“AI Research OS” orchestrator: idea → review → experiment → paper → submission |
$250k ARR |
Who we sell to first
- Beachhead: PhD scholars & early-career faculty in agriculture, biology, medicine and computer science in Pakistan, GCC and South Asia (large, price-sensitive, underserved, fast to adopt).
- Cheque-writers: ORIC directors, Deans of Research, research-lab PIs, R&D heads in pharma/agri companies.
How we make money
| Free / Student |
$0 |
3 reviews/month, watermark, community support |
| Individual Pro |
$20–30 / month (regional pricing) |
Unlimited runs, exports, priority models |
| Lab |
$2k–5k / year |
10 seats, shared library, admin |
| Institution |
$20k–100k / year |
Unlimited seats, SSO, training, support, reports |
| Enterprise (pharma / agri R&D) |
$100k–500k / year |
On-prem/VPC, custom pipelines, SLAs |
Guiding principles (signed by all five)
- Traceability over eloquence — no citation without a source ID.
- Independent modules, shared patterns — copy one folder, it still runs.
- Structured first — Pydantic models for every decision.
- Teach while building — every file is readable by a student of “Python ka Chilla”.
- Human in the loop — agents draft, humans decide; especially for supervision and publication.
Council sign-off
| Dr. Ayesha |
✅ |
“Ship the Literature Review Agent first, and make PRISMA the backbone.” |
| Bilal |
✅ |
“Same pattern in all five: Plan → Retrieve → Structure → Critique → Write.” |
| Sara |
✅ |
“Institutions pay; individuals evangelise. Pilot with one ORIC.” |
| Prof. Khan |
✅ |
“Data privacy and academic-integrity notices in every output.” |
| Hamza |
✅ |
“One .venv, one key, one command per agent. Everything goes in the book.” |
Codanics — Python ka Chilla: Build AI Agents
- Course: https://codanics.com/courses/python-ka-chilla-build-ai-agents/
- Website: https://codanics.com · YouTube: @codanics
- WhatsApp channel: https://whatsapp.com/channel/0029Va7nRDq3QxRzGqaQvS3r
- Complete free course: https://www.youtube.com/watch?v=UE3o-7KRNXs&list=PLaQmdklCcX40