$ EXECUTION PLAN — DAYS 1–90

PHASE 0: FOUNDATION

"Pehle ghar ki neenv banao, phir deewarein"

Build the foundation first, then the walls.

90
DAYS
/
$5–10M
BUDGET
/
25–30
PEOPLE
/
1T+
TOKENS
> MAIN GOALS
01

90-DAY TIMELINE

MONTH 1
Setup & Hiring
Days 1–30
→
MONTH 2
Building Core
Days 31–60
→
MONTH 3
Integration & Prototype
Days 61–90
MONTH 1

Setup & Hiring

Days 1–30
WEEK 1 Days 1–7
Company & Legal Setup
DAY 1–2: Legal Foundation
  • Company register karo (LLC / Pvt Ltd)
  • IP (Intellectual Property) protection
  • Legal counsel hire karo
  • NDAs draft karo (employees ke liye)
  • Banking setup
DAY 3–4: Workspace Setup
  • Office space / Remote-first setup
  • GitHub Enterprise, Linear, Notion, Slack, Zoom
  • Email domain: nexus-ai.com
DAY 5–7: Compute Access
  • AWS + GCP accounts setup
  • CoreWeave account (cheapest H100s)
  • Lambda Labs account (research GPUs)
  • $50K compute credits secure karo
WEEK 2 Days 8–14
Core Team Hiring
TIER 1 — HIRE IMMEDIATELY
CTO / VP Engineering
Google Brain / Meta AI / OpenAI background
$400–600K + equity
Lead Research Scientist
PhD ML/NLP · Transformer architecture
$350–500K + equity
Lead ML Engineer
DeepSpeed, Megatron-LM, NCCL
$300–450K + equity
Lead Data Engineer
Spark, Airflow, distributed systems
$250–350K + equity
Safety Researcher
RLHF, Constitutional AI, red-teaming
$300–450K + equity
TIER 2 — BY MONTH 2
  • 3× Senior ML Engineers
  • 2× Senior Backend Engineers (Python/FastAPI)
  • 1× DevOps / Infrastructure Engineer
  • 1× Senior Frontend Engineer
  • 2× Research Scientists
  • 3× Data Engineers
  • 1× Product Manager
TIER 3 — BY MONTH 3
  • 5× ML Engineers
  • 2× Backend Engineers
  • 1× Security Engineer
  • 3× Data Annotators / Curators
TOTAL: 25–30 PEOPLE
WEEK 3 Days 15–21
Development Environment
$ GitHub Org: nexus-ai
📁 nexus-ai/core main model code
📁 nexus-ai/training training pipeline
📁 nexus-ai/data data pipeline
📁 nexus-ai/inference inference engine
📁 nexus-ai/api API layer
📁 nexus-ai/frontend web app
📁 nexus-ai/safety safety systems
📁 nexus-ai/infra infrastructure
📁 nexus-ai/research experiments
Code Standards
  • Python style: PEP 8 + Ruff linter
  • Type hints: Required for all functions
  • Testing: pytest, minimum 80% coverage
  • Code review: Minimum 2 approvals
  • Commit messages: Conventional Commits
WEEK 4 Days 22–30
Research Kickoff
Architecture Decisions
Attention Type
MHA vs GQA vs MQA
→ GQA
best speed-quality tradeoff
Positional Encoding
RoPE vs ALiBi vs Learned
→ RoPE
extensible context
Activation Function
ReLU vs GELU vs SwiGLU
→ SwiGLU
best performance
Normalization
LayerNorm vs RMSNorm
→ RMSNorm
faster, simpler
Expert System
Dense vs MoE
→ MoE / Dense
MoE large, Dense small
MONTH 2

Building Core Components

Days 31–60
WEEK 5–6 Days 31–45
Core Infrastructure
☸ Kubernetes Cluster
  • CPU pool: 10× c5.4xlarge (API, services)
  • GPU pool: 4× p4d.24xlarge (8×A100 each)
  • Storage pool: High IOPS SSDs
  • NVIDIA GPU Operator + Cert-Manager
  • Prometheus Stack monitoring
🗄 Database Stack
  • PostgreSQL 16 (Primary + 2 read replicas)
  • Redis 7 Cluster (sessions, rate limiting, cache)
  • Milvus Vector DB (RAG embeddings)
  • MinIO S3-compatible storage
  • Delta Lake on MinIO
📊 Monitoring Stack
  • Prometheus (metrics)
  • Grafana (dashboards)
  • Loki (logs)
  • Jaeger (tracing)
  • AlertManager → Slack
✅ DELIVERABLE: Full dev infrastructure running
WEEK 6–7 Days 40–52
Data Pipeline v1
1
COLLECTION
CommonCrawl, Wikipedia, StackOverflow, GitHub, ArXiv, Urdu/Hindi sites
→
2
NORMALIZE
HTML→text, PDF→text, Unicode normalization
→
3
QUALITY FILTER
Length filter, perplexity scoring, heuristics
→
4
DEDUP
MD5 exact + MinHash LSH near-dedup
→
5
SAFETY FILTER
Toxicity, CSAM, PII detection & removal
→
6
CLASSIFY
Science, News, Code, Books, Math, General
Existing Datasets:
The Pile (800GB) C4 RedPajama Dolma The Stack mC4 CulturaX
🎯 TARGET: 1T+ tokens processed by end of Phase 0
WEEK 7–8 Days 45–60
Model Architecture v1
TOKENIZER
Priority: HIGHEST
  • BPE via SentencePiece
  • Vocabulary: 128,000 tokens
  • Byte fallback: True
  • Training: 3–5 days on CPU cluster
English
40%
Urdu ⭐
10%
Hindi
8%
Arabic
8%
Code
10%
TRANSFORMER STACK
NexusForCausalLM
embed_tokens
↓
N × NexusDecoderLayer
RMSNorm → GQA (Flash Attn 2) → Residual
RMSNorm → SwiGLU FFN → Residual
↓
Final RMSNorm
↓
lm_head (tied weights)
MODEL CONFIGS
nexus-tiny
Testing
512 hidden 8 layers 8 heads 8K ctx
nexus-small
Phase 0 Prototype
2048 hidden 24 layers 16 heads 32K ctx
MONTH 3

Integration & Prototype

Days 61–90
WEEK 9–10 Days 61–75
Training Pipeline v1
NEXUS-TINY
Params 100M
Data 10B tokens
GPUs 4× A100
Time ~2 days
Goal Validate pipeline
NEXUS-SMALL
Params 1.5B
Data 100B tokens
GPUs 8× A100
Time 7–10 days
Goal First real prototype
Training Config
  • Mixed precision: BF16
  • Optimizer: AdamW + cosine decay
  • Warmup: 2000 steps → cosine to 10% peak
  • Gradient clipping: 1.0
  • DDP → DeepSpeed ZeRO-3 upgrade path
WEEK 11–12 Days 76–90
Basic API & Interface
POST /v1/chat/completions OpenAI-compatible · SSE streaming
GET /v1/models List available models
GET /health Health check
POST /v1/completions Text completion
Frontend (Internal Demo)
  • Next.js chat interface
  • Streaming response display
  • Markdown + code syntax highlighting
  • Basic conversation history
Evaluation Framework v1
  • HellaSwag (few-shot)
  • MMLU (5-shot) — target: >40% for 1.5B
  • TriviaQA
02

BUDGET BREAKDOWN

SALARIES
$2.3M
10 Senior researchers/engineers × $40K/mo $1.2M
15 Mid-level engineers × $25K/mo $1.1M
GPU COMPUTE
$225K
8× A100 × 3mo × $10/hr × 24hr $170K
Nexus-Tiny training $5K
Nexus-Small training $50K
CLOUD INFRA
$150K
AWS/GCP (K8s, storage, networking) $50K/mo
3 months total $150K
TOOLS & SERVICES
$120K
GitHub Enterprise$5K
W&B (ML tracking)$10K
Existing datasets$100K
LEGAL & OPS
$350K
Legal fees$50K
Recruiting fees$200K
Office setup$100K
BUFFER
$500K
Unexpected costs$500K
TOTAL PHASE 0
$3.5–4M conservative
$5M with buffer & contingency
03

SUCCESS METRICS

TEAM
25+
people hired by Day 90
5
critical roles filled by Day 30
DATA
1T+
tokens in pipeline
50B+
Urdu tokens collected
>0.7
data quality score average
MODEL
<5
tokens/word for Urdu avg
<2.0
Nexus-Tiny training loss
>40%
MMLU 5-shot for 1.5B model
INFRASTRUCTURE
<1s
API latency for first token
>99%
uptime for dev environment
10K+
tokens/sec/GPU throughput
> PHASE 0 → PHASE 1 TRANSITION CRITERIA
☐ Team of 25+ people fully onboarded
☐ 1T+ tokens in data pipeline
☐ Tokenizer trained and validated
☐ Architecture code reviewed & tested
☐ Nexus-Small (1.5B) prototype working
☐ Training infrastructure proven (multi-GPU)
☐ Basic API serving requests
☐ CI/CD pipeline stable
☐ Phase 1 plan approved by leadership
☐ Phase 1 funding secured ($15–30M)
IF ALL ✅ → START PHASE 1!
Phase 1: Data Scale + Train 7B Model + Full Platform · 3–5 months · $15–30M
04

TEAM DAILY RHYTHM

09:00
Daily Standup
15 min · Yesterday / Today / Blockers
09:15
Deep Work Block 1
3 hours · Core development / research
12:15
Lunch Break
45 min
13:00
Deep Work Block 2
3 hours · Core development / research
16:00
Code Reviews & Collaboration
1 hour
17:00
Research Paper Presentation
Fridays only
MON Sprint planning
WED Technical deep-dive
FRI Research presentation + demo
BI-WEEKLY All-hands meeting