Confidential — Seed Round

The AI that
learns from every run.

NightShift is an autonomous problem solver with persistent memory.
Every run writes to a Knowledge Base. Run 50 is fundamentally different from Run 1.

665tests passing
70source files
3msKB query latency
9learning pillars
Seeking $1.5M seed  •  18-month runway
01 / 09
The Problem

Every AI coding tool
starts from zero. Every time.

Devin costs $9/hour and forgets everything when the session ends. Cursor doesn't know what worked yesterday. Copilot has no memory of your codebase patterns. Every run is Run 1.

$9/hr
Devin's base cost
Each session starts from scratch. Every failed strategy must be re-discovered. You pay to repeat the same mistakes.
0%
Knowledge carried forward
No tool on the market retains strategic knowledge across sessions. Statelessness is the bottleneck — not capability.
1x
Run 1 = Run 100
No learning curve. The AI doesn't get better at your problems. You bear all the adaptation cost yourself.
THE INSIGHT
The bottleneck in AI tools is not capability — it's memory. Teams are paying for the same discoveries over and over.
02 / 09
The Solution

NightShift remembers.
9 pillars of self-learning.

Problem in. Solution out. Code, research, analysis — any task. Every run writes to a persistent Knowledge Base. Every failure teaches. Every success compounds.

Pillar 1
Knowledge Base
LanceDB + ModernBERT. 3ms hybrid search. Global + local tiers.
Pillar 2
Agent Resources
Team patterns scored with UCB1. Mutate, compete, evolve.
Pillar 3
Auditor + Investor
Oversight pair. Prevent safe-play loops. Push explore/exploit signals.
Pillar 4
Evaluator
Honest judge. Git diff as truth. Spawns sub-evaluators for complex output.
Pillar 5
Unified Engine
One code path. All strategies. 13 features in a single engine.
Pillar 6
File Monitoring
status.json, events.jsonl, inbox. Any tool integrates. No API needed.
Pillar 7
Self-Learning
Episodic memory. UCB1 strategy scoring. KB consolidation after every run.
Pillar 8
Exploration
Investor prevents convergence. High uncertainty → bold moves.
Pillar 9
External Integration
inject, stop, status. User signals enter via same path as investor signals.
Capability Devin / Cursor / Copilot NightShift
Learns between runs No LanceDB KB — 3ms queries
Strategy from experience No AR patterns — UCB1 scoring
Non-code tasks No Research, analysis, SWOT
Smart retry on failure Full restart Blame-driven — only failed agents re-run
User intervention mid-run No Auditor inbox — inject anytime
Exploration drive No Investor agent — prevents safe-play loops
03 / 09
Market Opportunity

The memory layer
for agentic AI.

We're not competing with coding assistants. We're building the infrastructure layer that makes all agents better — starting with the developer market.

$10.9B
AI Agents market in 2026
Growing to $183B by 2033 (CAGR 49.6%) — Grand View Research
$14.3B
Developer tools market (2024)
18% YoY growth — IDC 2024
$8.1B
AI code tools market (2025)
Growing to $127B by 2032 (CAGR 48.1%) — MarketsandMarkets

Our position

Memory Layer. Every agentic AI system needs to learn from experience. NightShift is the first product that solves this for autonomous agents — not for human users, but for the agents themselves. This is a distinct, defensible position with no direct incumbent.

Go-to-market

Open source first. AGPL-3.0 drives distribution. Developers discover, self-host, trust. Cloud tier captures power users. Shared Knowledge Base becomes network effect — each contributor makes everyone's agents smarter.

Why now

Claude 3.5 / GPT-4o — capability is solved. Memory is the next bottleneck.
LanceDB + ModernBERT — infrastructure just became fast enough (3ms) to use mid-run.
Market proof — Cognition (Devin) raised $400M at $10.2B valuation, Sept 2025. $73M ARR. Category is real.
04 / 09
Business Model

Usage-based pricing.
Network effect via shared KB.

Self-Hosted

Free
AGPL-3.0. Forever.
  • ✓ Full NightShift engine
  • ✓ Local KB + AR patterns
  • ✓ Your own LLM API keys
  • ✓ All 9 pillars
GROWTH DRIVER

Cloud Pro

Usage-based
$0.003/KB query + $0.05/run
  • ✓ Hosted KB (no local setup)
  • ✓ LLM API included
  • ✓ Dashboard + history
  • ✓ Community KB access
  • ✓ Priority support
Typical user: 50 runs/mo, 200 queries/run → ~$12.50/mo

Enterprise

Custom
Volume + SLA
  • ✓ Private shared KB per team
  • ✓ SSO + audit logs
  • ✓ BYOK (encryption)
  • ✓ SLA + dedicated support
  • ✓ On-premise deployment

The Network Effect

Every user who opts into the Community KB contributes anonymized strategies, patterns, and domain knowledge. The more users contribute, the smarter every agent gets. This creates a compounding moat: a competitor starting fresh has no accumulated knowledge. A community with 10,000 runs of shared patterns is an infrastructure asset, not a feature.
Unit Economics
KB query (revenue)$0.003
KB query (LLM embedding cost)-$0.0004
Margin per query87%
Path to Breakeven
Monthly infra cost (1-person)~$3,000
Avg revenue per paying user~$12.50/mo
Breakeven at300 users
05 / 09
Traction

Working product.
Not a prototype.

665
Tests passing
70
Source files
17.5K
Lines of code
157
Git commits

What's shipped

  • ✓  Full 9-pillar engine — one unified code path
  • ✓  LanceDB KB with ModernBERT embeddings
  • ✓  AR patterns with UCB1 scoring + persistence
  • ✓  Auditor (mailbox, anomaly detection, diagnosis)
  • ✓  Investor (explore/exploit/deliver/reboot signals)
  • ✓  Evaluator with sub-evaluator spawning
  • ✓  Blame-driven retry (only failed nodes re-run)
  • ✓  File monitoring (status.json, events.jsonl, inbox)
  • ✓  Dashboard (standalone HTML, reads status.json)
  • ✓  CLI: solve, run, status, inject, stop

What's next (roadmap)

  • →  Server mode — HTTP API for remote problem submission
  • →  Investor Center — cross-project knowledge aggregation
  • →  Public Knowledge Base — opt-in community sharing
  • →  Dashboard polish — live WebSocket status updates
  • →  AR skill-level evolution — node performance optimization
  • →  Multi-user cloud — hosted version with per-user isolation
AGPL-3.0 license — open core model. Server mode = commercial moat. Companies that run NightShift as a service must open-source modifications or purchase a commercial license.
06 / 09
Team

What we're building
and who we need.

Current: Solo founder. The 9-pillar architecture, 665 tests, and full engine were built by one person. The core technical foundation is complete. What's needed now is speed: scale the learning system, build the cloud infrastructure, and grow the developer community.
Hire #1 — Priority
CTO / Core ML Engineer
$180–220k + equity
Owns the learning system: KB architecture, AR evolution, UCB1, Librarian consolidation. PhD-adjacent or strong ML research background. Ideally has worked on embedding models or vector search.
Hire #2
Infrastructure Engineer
$150–190k + equity
Server mode, LanceDB at scale, Docker/K8s, multi-tenant security. Experience with long-running job systems (queue-based, not serverless). LanceDB + S3 backend knowledge is a strong plus.
Hire #3
DevRel / Community Lead
$120–150k + equity
Open source community building, docs, conference talks, content. Has previously grown a dev tool GitHub repo to 5K+ stars. Understands developer psychology — not a marketer, a developer who communicates.
Hire #4
Designer / Frontend
$100–130k + equity
Dashboard UX, landing page quality, docs design system. Understands developer aesthetic — not corporate/safe design. Proficient with data visualization (run history, learning curves, KB graph).
Advisors sought: someone who has scaled a developer tool open source project to 50K+ GitHub stars, and someone with enterprise sales experience in developer infrastructure (Datadog, Grafana, etc. sales background).
07 / 09
The Ask

$1.5M seed.
18 months to product-market fit.

Budget allocation

Team (4 hires × 18 months)
$900k
Infrastructure (LLM API, compute)
$200k
Community + Marketing
$100k
Legal (AGPL enforcement, TM)
$50k
Buffer / Contingency
$250k

18-Month milestones

Months 1–3
Open Source Launch
HN Show, 500+ GitHub stars, 50 active users, community Discord live.
Months 4–6
Cloud Beta
Server mode live. 20 paying users. $500/mo MRR. First case studies published.
Months 7–12
Growth + Community KB
300 paying users. $3k+ MRR. Community KB launched. YC application or Series A prep.
Months 13–18
Enterprise + Series A
First enterprise deals. 1,000 users. $10k+ MRR. Series A raise with demonstrated network effect.
WHY THIS ROUND WORKS
The hard technical work is done. The 9-pillar architecture exists, the tests pass, the learning loops work. This capital funds the team and infrastructure to take a working product to market — not to build an unproven prototype. The risk profile of this investment is materially lower than a typical pre-product seed.
08 / 09
The Opportunity

Every AI tool today is stateless.
We're building the memory layer.

The capability race in AI is largely over. The differentiation now is in what agents learn. NightShift is the first product built on that thesis, with a working implementation and the right architecture.

9
Interconnected learning pillars
AGPL
Open core commercial moat
3ms
KB query — fast enough to use mid-run
View Product Read Docs
NightShift  •  AGPL-3.0  •  nightshift_launch_package_v1.md for full diligence
09 / 09